the 13th international conference on neural information ... · conference on neural information...
TRANSCRIPT
The 13th International Conference on Neural Information Processing
October 3-6, 2006 Hong Kong
Table of Contents
Welcome Message 1 Message from the ICONIP2006 Co-Chairs 2
Organization 4 Organizers, Sponsors, & Advisors 5 ICONIP2006 Organizing Committee 6 ICONIP2006 Program Committee 7
Conference Information 10 Registration and Conference Venue 11 Social Events 12 Useful Telephone Numbers 14
Technical Program 15 Tutorials 16 Workshops 23 Plenary Speeches 24 Invited Talks 25 Session Identifiers 35 Oral Presentation Guidelines 36 Poster Presentation Guidelines 37 Session Summary 38 Wednesday Session A, 4 October 2006 48 Wednesday Session B, 4 October 2006 65 Wednesday Session C, 4 October 2006 80 Wednesday Poster Session I, 4 October 2006 94 Thursday Session A, 5 October 2006 107 Thursday Session B, 5 October 2006 123 Thursday Session C, 5 October 2006 139 Thursday Poster Session II, 5 October 2006 154 Friday Session A, 6 October 2006 169 Friday Session B, 6 October 2006 187 Friday Poster Session III, 6 October 2006 200
Session Chair Index 214 Author Index 215
1
Welcome Message
2
Message from the ICONIP2006 Co-Chairs On behalf of the organizing team, we welcome you to the Thirteenth International Conference on Neural Information Processing (ICONIP2006). Sponsored by the Asia Pacific Neural Network Assembly (APNNA), ICONIP is a premier conference in the Asian and Pacific regions with significant global impact. The past events were held in Seoul (1994), Beijing (1995), Hong Kong (1996), Dunedin (1997), Kitakyushu (1998), Perth (1999), Taejon (2000), Shanghai (2001), Singapore (2002), Istanbul (2003), Calcutta (2004), and Taipei (2005). Over the years, ICONIP has matured into a well-established series of international conference on neural information processing and related fields. Following the tradition, ICONIP2006 will provide an academic forum for the participants to disseminate their new research findings and discuss emerging areas of research. It will also create a stimulating environment for the participants to interact and exchange information on future challenges and opportunities of neural network research. ICONIP2006 received 1,175 submissions from about 2,000 authors in 42 countries and regions (Argentina, Australia, Austria, Bangladesh, Belgium, Brazil, Canada, China, Hong Kong, Macao, Taiwan, Colombia, Costa Rica, Croatia, Egypt, Finland, France, Germany, Greece, India, Iran, Ireland, Israel, Italy, Japan, South Korea, Malaysia, Mexico, New Zealand, Poland, Portugal, Qatar, Romania, Russian Federation, Singapore, South Africa, Spain, Sweden, Thailand, Turkey, United Kingdom, and United States of America) across six continents (Asia, Europe, North America, South America, Africa, and Oceania). Based on rigorous reviews by the program committee members and reviewers, 386 high-quality papers were selected for publication in the proceedings with the acceptance rate being less than 33%. The papers are organized in 22 cohesive sections covering all major topics of neural network research and development. In addition to the contributed papers, the ICONIP2006 technical program includes two plenary speeches by Shunichi Amari and Russell Eberhart. In addition, ICONIP2006 program includes invited talks by the leaders of technical co-sponsors such as Wlodzislaw Duch (President of European Neural Network Society), Shiro Usui (President of Japanese Neural Network Society), DeLiang Wang (President of International Neural Network Society), and Shoujue Wang (President of China Neural Networks Council). In addition, ICONIP2006 launched the APNNA Presidential Lecture Series with invited talks by past APNNA Presidents and K.C. Wong Distinguished Lecture Series with invited talks by eminent Chinese scholars. Furthermore, the program also includes six excellent tutorials, opened to all conference delegates to attend, by Amir Atiya, Russell Eberhart, Mahesan Niranjan, Alex Smola, Koji Tsuda, and Xuegong Zhang. Besides the regular sessions, ICONIP2006 also features seven special sessions focusing on some emerging topics. ICONIP2006 would not achieve its success without the generous contributions of many volunteers and organizations. ICONIP2006 organizers would like to express sincere thanks to APNNA for the sponsorship, to China Neural Networks Council, European Neural Network Society, IEEE Computational Intelligence Society, IEEE Hong Kong Section Computer Chapter, International Neural Network Society, and Japanese Neural Network Society for their technical co-sponsorship, to the Chinese University of Hong Kong for its financial and logistic supports, and to the K.C. Wong Education Foundation of Hong Kong for its financial support. The organizers would also like to thank the members of the Advisory Committee for their guidance, the members of the International
3
Program Committee and additional reviewers for reviewing the papers, and members of the Publications Committee for checking the accepted papers in a short period of time. Particularly, the organizers would like to thank the proceedings publisher, Springer, for publishing the proceedings in the prestigious series of Lecture Notes in Computer Science. Special mention must be made to a group of dedicated students and associates, Dr. Haixuan Yang, Zhenjiang Lin, Zenglin Xu, Xiang Peng, Po Shan Cheng, and Terence Wong, who worked tirelessly and relentlessly behind the scene to make the mission possible. There are still many more colleagues, associates, friends, and supporters who have helped us in immeasurable ways that we might have missed; we express our sincere thanks to them all. Last but not least, the organizers would like to thank all the speakers and authors for their active participation at ICONIP2006, which made it a great success. Jun Wang and Laiwan Chan, General Chairs Irwin King and DeLiang Wang, Program Chairs Man Wai Mak, Organizing Chair
4
Organization
5
Organizers, Sponsors, & Advisors Organizer The Chinese University of Hong Kong Sponsor Asia Pacific Neural Network Assembly Financial Co-Sponsor K.C. Wong Education Foundation of Hong Kong Technical Co-Sponsors IEEE Computational Intelligence Society International Neural Network Society European Neural Network Society Japanese Neural Network Society China Neural Networks Council IEEE Hong Kong Section Computer Chapter Honorary Chair and Co-Chair Lei Xu, Hong Kong Shunichi Amari, Japan Advisory Board Walter J. Freeman, USA Toshio Fukuda, Japan Kunihiko Fukushima, Japan Tom Gedeon, Australia Zhenya He, China Nik Kasabov, New Zealand Okyay Kaynak, Turkey Anthony Kuh, USA Sun-Yuan Kung, USA Soo-Young Lee, Korea Chin-Teng Lin, Taiwan Erkki Oja, Finland Nikhil R. Pal, India Marios M. Polycarpou, USA Shiro Usui, Japan Benjamin W. Wah, USA Lipo Wang, Singapore Shoujue Wang, China Paul J. Werbos, USA You-Shou Wu, China Donald C. Wunsch II, USA Xin Yao, UK Yixin Zhong, China Jacek M. Zurada, USA
6
ICONIP2006 Organizing Committee General Chair and Co-Chair Jun Wang, Hong Kong Laiwan Chan, Hong Kong Organizing Chair Man-Wai Mak, Hong Kong Finance and Registration Chair Kai-Pui Lam, Hong Kong Workshops and Tutorials Chair James Kwok, Hong Kong Publications and Special Sessions Chair and Co-Chair Frank H. Leung, Hong Kong Jianwei Zhang, Germany Publicity Chair and Co-Chairs Jeffrey Xu Yu, Hong Kong Chris C. Yang, Hong Kong Derong Liu, USA Wlodzislaw Duch, Poland Local Arrangements Chair and Co-Chair Andrew Chi-Sing Leung, Hong Kong Eric Yu, Hong Kong Secretary Haixuan Yang, Hong Kong
7
ICONIP2006 Program Committee
Program Chair and Co-Chair Irwin King, Hong Kong DeLiang Wang, USA Program Committee Shigeo Abe, Japan Peter Andras, UK Sabri Arik, Turkey Abdesselam Bouzerdoum, Australia Ke Chen, UK Liang Chen, Canada Luonan Chen, Japan Zheru Chi, Hong Kong Sung-Bae Cho, Korea Sungzoon Cho, Korea Seungjin Choi, Korea Andrzej Cichocki, Japan Chuangyin Dang, Hong Kong Wai-Keung Fung, Canada Takeshi Furuhashi, Japan Artur dAvila Garcez, UK Daniel W.C. Ho, Hong Kong Edward Ho, Hong Kong Sanqing Hu, USA Guang-Bin Huang, Singapore Kaizhu Huang, China Malik Magdon Ismail, USA Takashi Kanamaru, Japan James Kwok, Hong Kong James Lam, Hong Kong Kai-Pui Lam, Hong Kong Doheon Lee, Korea Minho Lee, Korea Andrew Leung, Hong Kong Frank Leung, Hong Kong Yangmin Li, Macau Xun Liang, China Yanchun Liang, China Xiaofeng Liao, China Chih-Jen Lin, Taiwan Xiuwen Liu, USA Bao-Liang Lu, China Wenlian Lu, China Jinwen Ma, China Man-Wai Mak, Hong Kong Sushmita Mitra, India Paul Pang, New Zealand Jagath C. Ra japakse, Singapore
8
Bertram Shi, Hong Kong Daming Shi, Singapore Michael Small, Hong Kong Michael Stiber, USA Ponnuthurai N. Suganthan, Singapore Fuchun Sun, China Ron Sun, USA Johan A.K. Suykens, Belgium Norikazu Takahashi, Japan Michel Verleysen, Belgium Si Wu, UK Chris Yang, Hong Kong Hujun Yin, UK Eric Yu, Hong Kong Jeffrey Yu, Hong Kong Gerson Zaverucha, Brazil Byoung-Tak Zhang, Korea Liqing Zhang, China Reviewers Shotaro Akaho Toshio Akimitsu Damminda Alahakoon Aimee Betker Charles Brown Gavin Brown Jianting Cao Jinde Cao Hyi-Taek Ceong Pat Chan Samuel Chan Aiyou Chen Hongjun Chen Lihui Chen Shu-Heng Chen Xue-Wen Chen Chong-Ho Choi Jin-Young Choi M.H. Chu Sven Crone Bruce Curry Rohit Dhawan Deniz Erdogmus Ken Ferens Robert Fildes Tetsuo Furukawa John Q. Gan Kosuke Hamaguchi Yangbo He Steven Hoi
Pingkui Hou Zeng-Guang Hou Justin Huang Ya-Chi Huang Kunhuang Huarng Arthur Hsu Kazushi Ikeda Masumi Ishikawa Jaeseung Jeong Liu Ju Christian Jutten Mahmoud Kaboudan Sotaro Kawata Dae-Won Kim Dong-Hwa Kim Cleve Ku Shuichi Kurogi Cherry Lam Stanley Lam Toby Lam Hyoung-Joo Lee Raymond Lee Yuh-Jye Lee Chi-Hong Leung Bresley Lim Heui-Seok Lim Hsuan-Tien Lin Wei Lin Wilfred Lin Rujie Liu
9
Xiuxin Liu Xiwei Liu Zhi-Yong Liu Hongtao Lu Xuerong Mao Naoki Masuda Yicong Meng Zhiqing Meng Yutaka Nakamura Nicolas Navet Raymond Ng Rock Ng Edith Ngai Minh-Nhut Nguyen Kyosuke Nishida Yugang Niu YewSoon Ong Neyir Ozcan Keeneth Pao Ju H. Park Mario Pavone Renzo Perfetti Dinh-Tuan Pham Tu-Minh Phuong Libin Rong Akihiro Sato Xizhong Shen Jinhua Sheng Qiang Sheng Xizhi Shi Noritaka Shigei Hyunjung Shin Vimal Singh Vladimir Spinko Robert Stahlbock Hiromichi Suetant Jun Sun
Yanfeng Sun Takashi Takenouchi Yin Tang Thomas Trappenberg Chueh-Yung Tsao Satoki Uchiyama Feng Wan Dan Wang Rubin Wang Ruiqi Wang Yong Wang Hua Wen Michael K.Y. Wong Chunguo Wu Guoding Wu Qingxiang Wu Wei Wu Cheng Xiang Botong Xu Xu Xu Lin Yan Shaoze Yan Simon X. Yang Michael Yiu Junichiro Yoshimoto Enzhe Yu Fenghua Yuan Huaguang Zhang Jianyu Zhang Kun Zhang Liqing Zhang Peter G. Zhang Ya Zhang Ding-Xuan Zhou Jian Zhou Jin Zhou Jianke Zhu
10
Conference Information
11
Registration and Conference Venue
Full Registration Full registration includes tutorials, all technical sessions, CD proceedings, all coffee breaks, the conference reception, banquet, and farewell party. Student Registration Student registration includes tutorials, all technical sessions, all coffee breaks, the conference reception, and farewell party. Registration and Help Desk The Registration and Help Desk is outside Room 401 in the Foyer. The desk will be open from 8:30 am to 30 minutes after the last session during conference period for registration and inquiries. Extra tickets for social events may be purchased from the registration desk. Break Locations Coffee will be available in the Convention Foyer area during the coffee break time. There will be one in the morning and one in the afternoon. On Friday, October 6, there will be a Farewell Party in conjunction with the Poster Session III. Meeting Rooms Room 401 is used for Plenary and Invited Talks. Rooms 402 to 408 are used for parallel technical sessions. Room 409 is used as the Computer Room. Room 410 is used by the Organizing Committee. It can be used at a meeting room. Please contact the Registration and Help Desk for information on using Room 410. Internet Access A wireless network is provided throughout Rooms 401 to 410. Cable access and fixed terminals are available in Room 409 from 12:00 pm, Tuesday, October 3 to 4:30 pm, Friday, October 6. Message Board & Table There is a message board in the Foyer area for posting news, meeting announcements, and general information by the participants and also the Organizers. There will also be a table in the Foyer area for leaving CFPs, journals, papers, etc. Exhibition Two publishers, Springer and World Scientific, will take part in the book exhibition of the conference at the Foyer Area of Room 401, HKCEC.
12
Social Events
Opening Ceremony and Welcome Reception Information Venue: Room 401 Date: 3 October 2006 (Tuesday) Time: 6:00 p.m. Banquet Information Venue: Tao Yuan Restaurant, 3/F, Great Eagle Centre, 23 Harbour Road, Wanchai, Hong
Kong. Date: 5 October 2006 (Thursday) Time: 6:30 p.m. Banquet speaker: Ir. S. W. Cheung, Hong Kong Science and Technology Parks
Corporation (HKSTP) Topic: Innovation and Technology Development in Hong Kong (Tentative) 10 persons per table Farewell Party Information Venue: Foyer of Room 401 Date: 6 October 2006 (Friday) Time: 3:00 p.m. to 4:30 p.m. Lunch Information Many restaurants are located around the conference venue. Here are some of them inside the HKCEC:
• Golden Bauhinia (Cantonese cuisine) Tel: (852) 2582 7728 • Congress Restaurant (International cuisine) Tel: (852) 2582 7250 • Harbour Kitchen (Authentic Hong Kong flavours) Tel: (852) 2582 7241
You can also find various Chinese & Western restaurants inside the Great Eagle Centre, Harbour Centre, Sun Hung Kai Centre, and China Resources Building. All of them can be reached within 10 minutes walk. Around the HKCEC, you can find tea restaurants, fastfood restaurants and takeaways for budget lunch. Most of them can be found inside the Sun Hung Kai Centre and the China Resources Building.
13
14
Useful Telephone Numbers
Hong Kong International Dialing Code: 852 Directory Enquiries: 1081 Emergency Service (Police, Fire, Ambulance): 999 Hong Kong Tourism Board Visitor Hotline: 2508 1234 General Police Enquiries: 2527 7177 Hong Kong International Airport, English (24 hours): 2181 0000 Hong Kong Immigration Department (24 hours): 2824 6111
15
Technical Program
16
Tutorials
Morning Session, Tuesday, October 3, 2006 TU-A402, Machine Learning in High-Throughput Genomics and Proteomics Xuegong Zhang, Tsinghua University, China TU-A404, Kernel Methods for Estimation and Data Analysis Alex Smola, National ICT Australia / ANU, Australia TU-A406, Machine Learning Models for Stock Market Prediction and Trading System Design Amir Atiya, Cairo University, Egypt Afternoon Session, Tuesday, October 3, 2006 TU-B402, Kernel Methods for Structural Data in Computational Biology Koji Tsuda, MPI for Biological Cybernetics, Germany TU-B404, Introduction to Swarm Intelligence Russell Eberhart, Indiana University Purdue University Indianapolis, USA TU-B406, Sequential Inference using Particle Filters Mahesan Niranjan, University of Sheffield, UK
17
TU-A402, Xuegong Zhang Date: Oct 3 Time: 9:00-12:15 Venue: 402-403 Title: Machine Learning in High-Throughput Genomics and Proteomics Abstract High-throughput genomics and proteomics data have been a major source of information in the current systems biology investigations. Machine learning methods like support vector machines (SVMs), neural networks, dimension reduction etc. have been playing an active role the analysis and mining of these data, composing one of the major efforts in current bioinformatics research. A typical scenario is using gene expression data obtained with DNA microarrays or proteomics data obtained with mass-spectrometry for classifying the samples (e.g. normal vs. cancer, subtypes of the cancer, etc) and discovering the relevant genes behind the classification. This tutorial will provide a systematic and in-depth overview of this field, covering from the biological background and major issues to be studied to state-of-the-art approaches and representative application examples, as well as some common pitfalls, open questions and challenges. The focus will be on clustering, classification, gene/biomarker selection and the proper assessment of the results achieved with machine-learning approaches. Biography Xuegong Zhang received his BS degree in Industrial Automation at Tsinghua University in 1989, and his Ph.D. degree in Pattern Recognition and Intelligent Systems at Tsinghua University in 1994. He joined the faculty of Tsinghua University, Department of Automation as an Assistant Professor in 1994, Associate Professor in 1996 and Full Professor in 2002. During 2001-2002, Dr. Zhang worked at Harvard School of Public Health as a visiting scientist. He is now the director of the Bioinformatics Division of TNLIST (Tsinghua National Laboratory for Information Science and Technology), Professor of Pattern Recognition and Bioinformatics of the Department of Automation, Tsinghua University. His current research interests include the analysis of high-throughput biological data with machine learning methods, computational analysis of alternative splicing and microRNAs, computational analysis of haplotype blocks and meiotic recombination hotspots, and statistical learning methodology.
18
TU-A404, Alex Smola Date: Oct 3 Time: 9:00-12:15 Venue: 404-405 Title: Kernel Methods for Estimation and Data Analysis Abstract In this course I will discuss modern kernel methods for classification, regression and sequence annotation. The work is based on exponential families and graphical models. Using these tools one can obtain a unified and simple approach to estimation. In the second part of the tutorial I will discuss how nonparametric methods can be used for unsupervised learning, such as feature selection, independent component analysis, database merging and schema matching, and sample bias correction. The resulting algorithms are extremely simple to implement, yet they are better than the best currently available methods in statistics.
19
TU-A406, Amir Atiya Date: Oct 3 Time: 9:00-12:15 Venue: 406-407 Title: Machine Learning Models for Stock Market Prediction and Trading System
Design Abstract In this tutorial I will present a short introduction into the various issues involved in predicting and designing trading systems for the stock market. The main emphasis will be on two issues: reviewing the workings of the stock market and trading systems, and reviewing ways how to apply neural networks and other machine learning models to these problems. The starting point is to have a learning model that predicts stock movements well. However, this is only a small part of the whole problem. Many other issues have to be taken into account when building a successful system. We will also review some of these issues. In particular, we will consider the following issues:
1. Types of trading systems: trend-following/contrarian/value; categories of systems: arbitrage, sector switching, intraday, etc.
2. Inputs/indicators: technical: moving averages, patterns; fundamental: financial statement data, earnings surprises; techno-fundamental: insider sales, short ratio; market wide indicators: A/D ratio; economic indicators.
3. Learning models: neural networks, SVM, kernel regression, k-nearest neighbor, mixtures of experts, etc.
4. Design considerations: overfitting, data snooping, test of the null hypothesis of whether being profitable, survival bias, forward looking, market regime switch.
5. Other considerations: transactions costs, exit stops and profit targets, order types, etc. 6. Performance and risk measures: Sharpe ratio, maximum drawdown, Sterling ratio, etc. 7. Brief review of risk management and portfolio optimization.
Biography Amir Atiya has obtained his Ph.D. from Caltech, and has been active in research in the areas of neural networks, machine learning, computational methods, and computational finance. He has 13 years experience in trading system design. Currently he is an Associate Professor at Cairo University. He held a Visiting Associate position at Caltech from 1997-2001, and had research positions in several financial firms, such as Qantxx, Tradelink, Simplex Technology, Countrywide, and Dunn Capital Management. He has been active in the academic community as well, as a member of the organization committee for the yearly Computational Finance Conference, Neural Networks in the Capital Markets Conference, and as program co-chair for IEEE Conference on Computational Intelligence for Financial Engineering (CIFER’2003). He has been a Guest Editor of the Special Issue of IEEE Transactions Neural Networks on “Neural Networks in Financial Engineering”, that appeared in July 2001. He got several awards, including the INNS Young Investigator Award in 1996 and the Kuwait Prize in 2005. Currently, he is an associate editor for IEEE Transactions Neural Networks. He has published close to 100 papers on the topics of neural networks, statistical learning theory and applications of these.
20
TU-B402, Koji Tsuda Date: Oct 3 Time: 14:00-17:15 Venue: 402-403 Title: Kernel Methods for Structural Data in Computational Biology Abstract 0. Basics of Kernel Methods (30min)
• Definition of Kernels • Support Vector Machines • Kernel PCA • Kernel Combinations
1. Genome Sequence Analysis (45min) • Sequence Classification Problems
o Splice sites o Promotors o Translation Initiation Sites o etc.
• Kernels for Sequences o Spectrum kernels o Weighted Degree kernels o Marginalized kernels o etc.
• Actual Applications o Reports published results
2. Graph Classification: Chemical Compounds and Proteins (45min) • Graph Classification Problems
o Chemical Compounds: Prediction of Toxicity, Mutagenecity, etc. o Protein 3D structures o RNA secondary structures
• Graph kernels o Random walk-based definition o Efficient Computation
• Actual Applications o Reports published results
3. Biological Network Analysis (45min) • Various Biological Networks
o Metabolic network o Protein interaction network o Gene regulatory network o etc.
• Protein Classification based on Networks o Diffusion kernels o Laplacian-based Methods
• Network Inference using Kernels o Supervised Network Inference o Kernel CCA o Combination of multiple data
4. Wrapping up (15min)
21
TU-B404, Russell Eberhart Date: Oct 3 Time: 14:00-17:15 Venue: 404-405 Title: Introduction to Swarm Intelligence Abstract The tutorial provides an overview of swarm intelligence, with a focus on particle swarm optimization. The tutorial begins with a brief review of dynamic social impact theory, and of evolutionary computation. This is followed by an introduction to particle swarm optimization (PSO) and the evolution of PSO concepts and paradigms. PSO features and processes are then discussed, followed by a review of PSO applications. Evolving neural networks with PSO is described. Finally, tracking and optimizing dynamic systems with PSO are reviewed, including a demonstration.
22
TU-B406, Mahesan Niranjan Date: Oct 3 Time: 14:00-17:15 Venue: 406-407 Title: Sequential Inference using Particle Filters Abstract Many modern signal processing problems involve systems that are nonlinear and nonstationary. Data-driven models that are based on powerful function approximation methods such as neural networks have been applied with demonstrable success to these problems. Nonstationarity imposes a particular difficulty in these settings because regularisation techniques such as cross validation can be inapplicable. This tutorial will address sequential estimation techniques that are useful in nonlinear and nonstationary environments. It will use a Bayesian dynamical systems approach and will introduce concepts and algorithms involving the extended Kalman filter (EKF) and powerful variants of it. Starting from the EKF, we will review more recent developments in sequential Markov Chain Monte Carlo (Particle Filters), and explore their application in a number of practical examples taken from speech signal and speech processing. Biography Mahesan Niranjan received his BSc from the University of Peradeniya, Sri Lanka (1982) and MEE from Eindhoven, The Netherlands (1985), both in Electronics Engineering. His PhD was from the University of Cambridge, England (1990). After eight years as lecturer in the Cambridge University Engineering Department, he was appointed to a Professorship in Computer Science in The University of Sheffield, where also served as Head of the Department of Computer Science. In Sheffield he leads a research group in Machine Learning which includes three members of Faculty, three post-doctoral research assistants and ten PhD candidates. His research interests are in Nonlinear & Nonstationary Signal Processing and Statistical Pattern Recognition. He has worked on the algorithmic aspects as well as on a range of applications including Speech Processing, Medical Diagnostics and Computational Finance. More recently he is interested in problems in Computational Biology and works closely with a team of Developmental Biologists in Sheffield.
23
Workshops
Date: Oct 4 Time: 13:30-17:30 Venue: 401 Relationship between Z-curve and FFT Approaches for DNA Coding Sequence Classification N. F. Law, Hong Kong Polytechnic University 13:30 – 14:00 Biclustering Gene Expression Profiles by Alternately Sorting with Weighted Correlated Coefficient L.W. Chan, Chinese University of Hong Kong 14:00 – 14:30 Feature Selection and Analysis for Promoter Prediction Based on Relative Entropy and Information Measures H. Yan, City University of Hong Kong 14:30 – 15:00 Tea Break and Poster Session 15:00 – 16:30 Network Component Analysis C. Q. Chang, University of Hong Kong 16:30 – 17:00 Eukaryotic Protein Subcellular Localization Based on Local Pairwise Profile Alignment SVM M.W. Mak, Hong Kong Polytechnic University 17:00 – 17:30
24
Plenary Speeches Integration of Stochastic Models in Neural Inference Shun-ichi Amari, RIKEN Brain Science, Japan Wednesday, 9:00-10:00, Room 401 Abstract The brain needs to integrate various probability distributions for stochastic inference. When there are a number of probability distributions, how is the good way of integration? The weighted average (arithmetic mean) is one method, but another is the geometric mean. We generalize the idea of the mean, to give the alpha-mean, where alpha is a real number. We then define the alpha-family of probability distributions. This is related to information geometry. The characteristics of the alpha-integration are explained. We then define the generalized divergence measure (alpha-divergence), and show the alpha-integration is optimal under the criterion of minimizing the alpha-divergence. The idea is used to understand the population coding, sensory integration, and the models of mixture of experts or the product of the experts. The Bayesian predictive distribution is also generalized to the Bayesian alpha-predictive distribution. Swarm Intelligence and Extended Analog Computing Russell Eberhart, Indiana University Purdue University Indianapolis, USA Thursday, 9:00-10:00, Room 401 Abstract Recent developments in swarm intelligence and particle swarm optimization are presented. Evolving neural networks using swarm intelligence is discussed. An overview of the relatively new field of extended analog computing is presented, and the use of swarm intelligence to evolve configurations of extended analog computers is described. Configuration evolution will be demonstrated with an extended analog computer during the plenary presentation.
25
Invited Talks
Wednesday, October 4, 2006 WA401, An Auditory Scene Analysis Approach to Cocktail Party Processing DeLiang Wang, Oticon A/S and The Ohio State University, USA WA401, Towards comprehensive foundations of computational intelligence Włodzisław Duch, Nicolaus Copernicus University, Poland Thursday, October 5, 2006 TA401, Neuroinformatics Japan-node and Automatic Keyword Indexing for the Platform Shiro Usui, RIKEN, Brain Science Institute, Japan TA401, High-Dimensional Space Geometry Informatics and Its Applications Shoujue Wang, Chinese Academy of Sciences, China TB401, Connecting Missing Links between Neuroscience and Machine Perception: Feature Extraction, Binaural Processing, Top-Down Attention, and Audio-Visual Integration Soo-Young Lee, Korea Advanced Institute of Science and Technology, Korea TB401, Brain-, Gene-, and Quantum Inspired Connectionist Systems: Challenges and Opportunities Nikola Kasabov, Auckland University of Technology, Auckland, New Zealand TC401, The Value-Based Decision-Making, from Fruit Fly to Human Being Aike Guo, Chinese Academy of Sciences, China TC401, A hybrid intelligent optimal control method for the whole production line and applications Tianyou Chai, Northeastern University, China Friday, October 6, 2006 FA401, Mechanism Approach to Intelligence Research Yixin Zhong, Beijing University of Posts and Telecommunications, China FA401, Analyzing financial returns: factor models and GARCH effect Laiwan Chan, The Chinese University of Hong Kong, Hong Kong, China FB401, Score matching: a new alternative to MCMC for estimation of non-normalized statistical models Aapo Hyvarinen, University of Helsinki, Finland
26
WA401, An Auditory Scene Analysis Approach to Cocktail Party Processing (DeLiang Wang) The acoustic environment is typically composed of multiple simultaneous events. A remarkable achievement of the auditory system is its ability to disentangle the acoustic mixture and group the sound energy that originates from the same event or source. This process of auditory organization is referred to as auditory scene analysis. The cocktail party problem, or segregation of speech from interfering sounds, has proven to be extremely challenging computationally. In this talk I describe an auditory scene analysis approach to the cocktail party problem. Our model performs segmentation and grouping in a two-dimensional time-frequency representation that encodes proximity in frequency and time, periodicity, amplitude modulation, and onset/offset. The model has led to the state-of-the-art performance in speech segregation, and demonstrates the promise of the auditory scene analysis approach. WA401, Towards comprehensive foundations of computational intelligence (Włodzisław Duch) Computational intelligence (CI) at present is a mixture of pattern recognition, machine learning, bio-inspired optimization, control, neural and fuzzy knowledge modeling methods and applications. Grand challenges to CI and promising directions to solve them are outlined: the need for good foundations and proposals for such foundations are discussed, including heterogeneous systems, meta-learning to find simplest and most useful models, new goals of learning, and methods that go beyond pattern recognition problems. Prospects of scaling up intelligent systems to human level competence using neurocognitive inspirations are presented. Biography Wlodzislaw Duch heads the Department of Informatics, Nicolaus Copernicus University, Torun, Poland, and is a Visiting Professor at Nanyang Technological University, Singapore (2003-7). Ph.D. in quantum chemistry (1980), postdoc at USC, Los Angeles (1980-82), D.Sc. in applied math (1987); worked at University of Florida; Max-Planck-Institute, Munich, Germany, Kyushu Institute of Technology, Meiji and Rikkyo University in Japan, and several other institutions. He is on the editorial board of IEEE TNN, CPC, NIP-LR, Journal of Mind and Behavior, and 8 other journals; co-founder & scientific editor of the “Polish Cognitive Science” journal; president of the European NNS executive committee (2006-2008), member of IEEE NNS Technical committee; expert of the European Union science programs; published over 350 scientific and popular articles, 4 books, edited many others, his DuchSoft company makes GhostMiner software package marketed by Fujitsu. TA401, Neuroinformatics Japan-node and Automatic Keyword Indexing for the Platform (Shiro Usui) Neuroinformatics (NI) combine neuroscience and informatics research to develop and apply advanced tools and approaches essential for a major advancement in understanding the structure and function of the brain. We here present the world recent trends of NI and
27
present activities in Japan under the Japan-node. Since each NI platform needs keyword index tree for navigating the contents of the database, we here focus on a tool we developed to aid this activity. The results suggest that the algorithm is useful in automating the process of extracting highly potential keywords which significantly hastens the process of building the index tree. Biography Shiro USUI (PhD). Graduate from University of California, Berkeley in 1974. Fellow of IEEE (1994). He had been a Professor at Toyohashi University of Technology until 2003 and became the team leader of a new Lab for Neuroinformatics at the RIKEN Brain Science Institute. In 2005, he becomes the Vice-Director of Neuroinformatics Japan Center at the Institute. He is now the President of Japanese Neural Network Society and a Fellow of the Institute of Electronics, Information and Communication Engineers, Japan (2002). He had been the Project Leader of the Target oriented research and development in brain science under the Special Coordination Funds for Promoting Science and Technology by the MEXT: “Neuroinformatics Research in Vision” (1999–2004). TA401, High-Dimensional Space Geometry Informatics and Its Applications (Shoujue Wang) Biography Wang Shoujue is a native of Suzhou, Jiangsu Province. He was born in Shanghai and graduated from Tongji University in 1949. Wang served as a researcher at the Semiconductor Research Institute of the Chinese Academy of Sciences. In 1958, he successfully developed a germanium alloy diffusion high frequency transistor and put them into production. His invention was then applied to high-speed transistor computers. From 1959 to 1963, he created whole silicon planar technology and developed five kinds of silicon plane-type transistors. In the 1970s, he manufactured a large-scale IC mask plate by automatic plate making technology and later engaged in the research of new circuits. Wang then focused on multiple value and continuous logical circuit system research and put it into practical application. He was elected as an academician of the Chinese Academy of Sciences in 1980. TB401, Connecting Missing Links between Neuroscience and Machine Perception: Feature Extraction, Binaural Processing, Top-Down Attention, and Audio-Visual Integration (Soo-Young Lee) In this talk we will present several examples of the brain-inspired machine perception, which are possible by connecting the missing links between the neuroscience and information technology. The developed mathematical models of the human auditory pathway are integrated into a speech recognition system, of which 4 components are (1) the nonlinear feature extraction model from cochlea to auditory cortex, (2) the binaural
28
processing model at superior olivery complex, (3) the top-down attention model from higher brain to the cochlea, and (4) audio-visual integration based on the top-down attention. The unsupervised Independent Component Analysis shows that some auditory feature extraction and binaural processing mechanisms follow information theory with sparse representation. The ICA-based features resemble frequency-limited features extracted from the cochlea and also more complex time-frequency features from the inferior colliculus and auditory cortex. The latter includes onset/offset, multifrequency timbre, and frequency modulated features. The ICA-based binaural processing model utilizes the cochlear filterbank, and sound localization and blind signal separation are conducted at each filtered channel. Unlike popular Jeffreys’ model the developed model is applicable to convolved mixtures. Actually the sound localization based on Jeffrey’s model may be regarded as the starting point of the adaptive blind signal separation algorithm for speech enhancement. The coupling among frequency bands may also be incorporated by Independent Vector Analysis. The top-down attention model shows how the pre-acquired knowledge in our brain filters out irrelevant features or fills in missing features in the sensory data. It utilizes an attention filter similar to the “early filter” theory, but the attention filter coefficients are adapted based on the top-down signals from the perceived attention cue. Both the top-down attention and bottom-up binaural processing are combined into a single system for high-noisy cases. The top-down attention model is also applicable to audio-visual integration, and successfully explained the McGurk effect for incongruent audio and video data. Also, with this model the speech recognition performance in very noisy environments is greatly improved. Biography Soo-Young Lee received B.S., M.S., and Ph.D. degrees from Seoul National University in 1975, Korea Advanced Institute of Science in 1977, and Polytechnic Institute of New York in 1984, respectively. From 1977 to 1980 he worked for the Taihan Engineering Co., Seoul, Korea. From 1982 to 1985 he also worked for General Physics Corporation at Columbia, MD, USA. In early 1986 he joined the Department of Electrical Engineering, Korea Advanced Institute of Science and Technology, as an Assistant Professor and now is a Full Professor at the Department of BioSystems and also department of Electrical Engineering and Computer Science. In 1997 he established Brain Science Research Center, which is the main research organization for the Korean Brain Neuroinformatics Research Program. The research program is one of the Korean Brain Research Promotion Initiatives sponsored by Korean Ministry of Science and Technology from 1998 to 2008, and currently about 35 Ph.D. researchers have joined the research program from many Korean universities. He is a Past-President of Asia-Pacific Neural Network Assembly, and has contributed to International Conference on Neural Information Processing as Conference Chair (2000), Conference Vice Co-Chair (2003), and Program Co-Chair (1994, 2002). Dr. Lee is the Editor-in-Chief of the newly-established online/offline journal with double-blind review process, Neural Information Processing - Letters and
29
Reviews (www.nip-lr.info), and is on Editorial Board for 3 international journals. He received Leadership Award and Presidential Award from International Neural Network Society in 1994 and 2001, respectively, and APPNA Excellent Service Award in 2004. His research interests have resided in artificial brain, the human-like intelligent systems based on biological information processing mechanism in our brain. He has worked on the auditory models from the cochlea to the auditory cortex for noisy speech processing, information-theoretic binaural processing models for sound localization and speech enhancement, the unsupervised pro-active developmental models of human knowledge with multi-modal man-machine interactions, and the top-down selective attention models for superimposed pattern recognitions. His research scope covers the mathematical models, neuromorphic chips, and real-world applications. Also, he had recently extended his research into brain-computer interfaces with simultaneous fMRI and EEG measurements. TB401, Brain-, Gene-, and Quantum Inspired Connectionist Systems: Challenges and Opportunities (Nikola Kasabov) The talk discusses opportunities and challenges for the creation of artificial neural network (ANN) models inspired by principles at different levels of information processing in the brain * neuronal-, genetic-, and quantum, and mainly * the issues related to the integration of these principles into more powerful and accurate ANN models. A particular type of ANN, evolving connectionist systems (ECOS), is used to illustrate this approach. ECOS evolve their structure and functionality through continuous learning from data and facilitate data and knowledge integration and knowledge elucidation. ECOS gain inspiration from the evolving processes in the brain. Evolving spiking neural networks are presented too. With more genetic information becoming available now, it becomes possible to integrate the gene and the neuronal information into neuro-genetic models and to use them for a better understanding of complex brain processes. Further down in the information processing hierarchy, are the quantum processes. Quantum inspired ANN may help solve efficiently some hard computational problems. It may be possible to integrate quantum principles into brain-gene inspired ANN models for a faster and more accurate modeling. All the topics above are illustrated with some contemporary solutions, but many more open questions and challenges are raised and directions for further research outlined. Keywords: Artificial neural networks, Computational intelligence, Neuro-informatics, Bioinformatics, Evolving connectionist systems, Gene regulatory networks, Computational neurogenetic modeling, Quantum information processing. TC401, The Value-Based Decision-Making, from Fruit Fly to Human Being (Aike Guo) To explore the neural mechanisms of the value-based decision making under uncertainty is essential because it is a fundamental intelligent activity at every societal level (Hsuet al., 2005). The shared interest in understanding decision making has resulted in the emergence of a new interdisciplinary field of research, as ‘Neuroeconomics’, the goal of
30
which is to understand the neural correlates of our economic choice behavior (Sugrue et al., 2005). To make adaptive decisions, animals must evaluate the costs and benefits of available array of options. Dopamine neurons acting as ‘common reward currency’ are sensitive to the uncertainty of both the occurrence and the time of reward. Due to its sophisticated genetics, relatively simple anatomy, behavioral richness, and its remarkable molecular similarity to mammals, fruit flies Drosophila became the “Jack of all trades” in Life Science. We have discovered a value-based choice behavior among competing alternatives in Drosophila. Flies in the “paradoxical” situation resolve the “Color /Shape dilemma” by taking into account the relative salience of the current information during retrieval, but flies lacking mushroom bodies, a prominent part in the Drosophila brain involved in many behaviors, seem to have difficulties in resolving “conflicting” situations. Thus, Drosophila has greater cognitive processing abilities than they are usually given credit for (Tang & Guo, 2001). “The nose smells what the eye sees, and that happened in human crossmodal perception” (Gottfried and Dolan, 2003). We have shown that Drosophila is capable of enhancing the cross-modal (visual and olfactory) learning and memory. We found that conditioning with concurrent visual and olfactory cues reduced the threshold for unimodal memory retrieval. Furthermore, bi-modal preconditioning followed by unimodal conditioning with either a visual or an olfactory cue led to cross-modal memory transfer. Thus we have shown that the memory acquisition and subsequent retrieval of fruit flies are strongly influenced by input from multiple sensory modalities. Visual or olfactory stimuli which are too weak to elicit memory acquisition on their own can facilitate memory when presented concurrently (Guo & Guo, 2005). We quest for, does the value-based decision circuit in flies share similar circuit principle with that in primates? We will focus in near future on the reinforcement learning principles mediated by dopamine systems because converging evidences from primates and Drosophila fit well with a major role of dopamine system in decision-making, as well as a central role in guiding our behaviors and thoughts as well. Biography Prof. Aike Guo, Neuroscientist and Biophysicist, now serves as the research professor at the CAS Institute of Biophysics and Institute of Neuroscience, Shanghai Institutes for Biological Sciences. He obtained his doctor degree in Natural Sciences from Munich University, Germany in 1979. He is the Associate Director of ION. He was the chief scientist for the 973 Program (The National Basic Research Program) “Brain Development and Plasticity” (2000-2005) and he is now the chief scientist of the new 973-program “The Brain plasticity in structures and functions” (2006-2010). He became an academician of the Chinese Academy of Sciences in 2003. Dr. Guo’s current research interest is focusing on the molecular, cellular, and circuit integrative mechanisms underlying learning, memory and cognition-
31
like behaviour, e.g. value-based decision-making and cross-modal learning in fruit flies Drosophila. He is also interesting in the neurodegeneration, such as AD,PD and addiction. TC401, A hybrid intelligent optimal control method for the whole production line and applications (Tianyou Chai) With ever increased needs for an improved product quality, production efficiency, and cost in today’s globalized world market, advanced process control should not only realize the accuracy of each control loops, but also has the ability to achieve an optimization control of global production indices that are closely related to the improved product quality, enhanced production efficiency and reduced consumption. As a result, the optimal control for the global production indices has attracted an increased attention of various process industries. The optimal control of the global production indices requires an optimal combination of the production indices, technical indices and the operation of each control loops. In this paper, a hybrid intelligent control strategy is proposed for process industries. This new strategy consists of three control layers, namely the intelligent optimization of the global production indices, the intelligent optimal control of the technical indices and the intelligent process control. The intelligent optimization of the global production indices is composed of the setting model of the technical indices, the predictor of the global production indices, the feedback and prediction analysis adjustment models. The intelligent optimal control of the technique indices consists of the setpoints model of control loops, the prediction of technical indices, and the feedback and feedfoword regulators. The intelligent process control is then composed of normal decoupled PID controllers, decoupled nonlinear PID controllers with a neural network feedforword compensator for un-modeled dynamics and a switching mechanism. Such a control structure can automatically transfer the global production indices into a required number of setpoints for each control loops. Moreover, when the system is subjected to either operating-point changes or unexpected disturbances, setpoints of the control loops can be adaptively updated and the outputs of the control loops are made to follow the updated setpoints so that the global production indices can be controlled into their targeted ranges to realize the optimization control of the global production indices. The proposed method has been successfully applied to the largest hematite minerals processing factory in China, where remarkable social and economic benefits have been achieved. Such an industrial application has successfully demonstrated the performance of the proposed optimal control method which will therefore has a high potential for further and much wider applications. Biography Professor Tianyou Chai received the Ph.D. degree in 1985, became professor of Northeastern University, Liaoning Province, in 1988 and was appointed Ph.D. supervisor in 1990. He was elected a member of Chinese Academy of Engineering in 2003. He was member of IFAC Technical Board and chairman of IFAC Coordinating Committee on Manufacturing and Instrumentation during 1996-1999. He is a member of Chinese National Disciplinary Appraisal Group, vice-director of Committee of Experts of Advanced Manufacturing and Automation in National High-tech Program, director of the
32
National Engineering and Technology Research Center of Metallurgical Industry Automation at Northeastern University. Prof. Chai’s research interests include adaptive control, decoupling control, optimizing control and integrated automation of complex industrial processes. He has gained innovative research achievements in these fields. He has been principal investigator for more than 30 research projects from Chinese National Natural Science Foundation, Chinese National High-tech Program, Chinese National Key Technology Program, National High-tech Industrialization Priority Project and Key Automation Engineering of Enterprises. He has supervised nine postdoctoral fellows, more than fifty PhD students and one hundred master students. He has published more than two hundred papers in international journals and international conferences. He was invited to give plenary talk in ten international conferences, including 2005 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2004 IEEE Conference on Cybernetic and Intelligent Systems, 2006 IEEE International Conference on Networking, Sensing and Control. He is Editor-in-Chief of Control Engineering of China. And he served as General Chair of the 6th World Congress on Intelligent Control and Automation. He has twice received the Second-Grade Awards of National Science and Technology Progress of China, and 10 Top-Grade and First-Grade Awards of Science and Technology Progress from various Ministries of China and Liaoning Province. For his significant contributions to Chinese automatic control technology and industrial automation, he won 2002 Technological Science Progress Award from Ho Leung Ho Lee Foundation (one of the most prestigious scientific awards in China) and 2003 Science and Technology Honor Prize of Liaoning Province (the most prestigious scientific award in Liaoning proving, which is only awarded to two persons every 2 years). FA401, Mechanism Approach to Intelligence Research (Yixin Zhong) Neural networks, expert systems, and sensory-motor systems are three representatives of the major approaches - i.e., the structuralism approach, functionalism approach, and behaviorism approach - to intelligence research that have been widely employed in literature so far. A new approach, namely Mechanism Approach, an approach based on the core mechanism of intelligence formation, will be reported in the talk. It has been discovered that the core mechanism of intelligence formation in most cases can be expressed as transformations that transform information to knowledge and further to intelligent strategy. It has also been discovered that the three mentioned approaches can be harmoniously unified within the framework of mechanism approach. These discoveries may be of significance to the intelligence research. Biography Yixin Zhong is Professor from Beijing University of Posts and Telecommunications (BUPT), Beijing, China. He served as program co-chair of IJCNN’92 during which the idea of APNNA and ICONIP was formed. He had been Associate Editor to IEEE Transactions on Neural Networks (IEEE-TNN) from 1993-2005, Chairman of China Neural Networks Council (CNNC) from 2001 to 2005, and president of APNNA from 2001-2002. He has served as President of Chinese Association for Artificial Intelligence since 2001. The major areas of his research include neural networks, artificial
33
intelligence, information science, and natural language processing and understanding. He has contributed to the related areas 16 books and over 300 papers. FA401, Analyzing financial returns: factor models and GARCH effect (Laiwan Chan) Inter-correlation and volatility clustering are two major properties of financial returns. Volatility clustering (i.e. large changes in prices are often clustered together) reflects the temporal dependence in financial returns and is often successfully captured by the GARCH model. Inter-correlation reflects the spatial dependence in different return series, and factor models can be used to describe the relations among return series. We have applied the Independent Component Analysis (ICA), a popular statistical technique to extract the factors and form the independent factor model. By capturing and analyzing both inter-correlation and volatility clustering properties together, we could obtain a better understanding of the underlying characteristics of the financial data. In this talk, we will discuss the application of independent factor model in financial data analysis and show a few ways to combine independent factor model and the GARCH models to forecast the multivariate conditional covariance matrix of returns. Different criteria were exploited to extract the factors and their extensive comparisons were presented. Biography Prof. Lai-wan Chan received her BA, MA and PhD degrees in Engineering from the University of Cambridge. She is currently a Professor in the Computer Science and Engineering Department of the Chinese University of Hong Kong, and the Associate Dean (Education) in the Faculty of Engineering. Her research interest is in data mining, financial engineering, bioinformatics and artificial neural networks. In particular, the learning and modelling of the feedforward and the recurrent networks, and applications of neural networks in time series prediction, trading systems, portfolio management, image recognition, analysis of DNA microarrays, Cantonese speech recognition and multimedia database. Professor Chan has published more than 100 scientific papers in the above research areas. She was the organizing co-chair of the 1996 International Conference on Neural Information Processing (ICONIP96) and the general co-chair of ICONIP2006, the program chair/co-chair of the 1998 and 2000 International Conference on Intelligent Data Engineering and Learning (IDEAL) and the steering committee co-chair of IDEAL2003, 2004 and 2005. FB401, Score matching: a new alternative to MCMC for estimation of non-normalized statistical models (Aapo Hyvarinen) One often wants to estimate statistical models where the probability density function is known only up to a multiplicative normalization constant. Typically, one then has to resort to Markov Chain Monte Carlo methods, or approximations of the normalization constant. Here, we propose that such models can be estimated by minimizing the expected squared distance between the gradient (w.r.t. the data variable) of the log-
34
density given by the model and the gradient of the log-density of the observed data. While the estimation of the gradient of log-density function is, in principle, a very difficult non-parametric problem, we prove a surprising result that gives a simple formula for this objective function. The density function of the observed data does not appear in this formula, which simplifies to a sample average of a sum of some derivatives of the log-density given by the model. The method can be shown to provide consistent estimators. Furthermore, in the case of some exponential families, the solution can actually be obtained in closed form. In addition to its computational simplicity, the method can be shown to be statistically optimal for the purpose of Bayesian denoising of signals. Biography Aapo Hyvärinen studied undergraduate mathematics and statistics at the universities of Helsinki (Finland), Vienna (Austria), and Paris (France), and finally obtained a Ph.D. degree in Information Science at the Helsinki University of Technology in 1997. After further post-doctoral work at the Neural Network Research Center of the Helsinki University of Technology, he moved in 2003 to the Dept of Computer Science of the University of Helsinki. He has double affiliation with the Helsinki Institute for Information Technology, where he holds the position of Senior Research Scientist and leads a research group on neuroinformatics. He is the first author of the book “Independent Component Analysis”, and author or coauthor of more than 100 scientific articles. He was Co-Editor-in-Chief of International Journal of Neural Systems in 2001-2006, and he is currently Action Editor at Journal of Machine Learning Research and Neural Computation, as well as Contributing Faculty Member of Faculty of 1000 Biology. His research interests include unsupervised learning (in particular, generative statistical models), computational systems neuroscience, and the visual system.
35
Session Identifiers
The sessions are identified by the day, time, room, and paper sequence. The day is either [W]ednesday, [T]hursday, or [F]riday. The time is either [A] for the morning session, [B] for the early afternoon session, and [C] for the late afternoon session. Room is a number from 401 to 408. The last digit is the presentation sequence number, usually from 1 to 6. For example, the session identifier, “WA401-1” is the first paper in the Wednesday morning session in room 401. “TC408-3” is the third paper in the Thursday late afternoon session in room 408. For poster sessions, the identifiers are WP, TP, and FP for Wednesday Poster, Thursday Poster, and Friday Poster sessions, respectively. For inquiries, please contact Helpers throughout the conference ground or at the Registration Desk.
36
Oral Presentation Guidelines There are five parallel oral sessions during each period. Each oral paper is allocated 20 minutes, with 15 minutes for presentation and 5 minutes for questions and answers. Speakers should go to the session room at least 15 minutes before their session starts, introduce themselves to the session chairs and check their presentation material with the computer and audio-visual equipment. The computer in each session room can display MS PowerPoint and Adobe PDF files. Speakers can bring their presentation material on USB drives. If you use other digital storage devices not supported by the computer in the session room, please ask a conference helper to transfer the files.
37
Poster Presentation Guidelines
Poster Layout The poster board's orientation is vertical (portrait). It is 1 m (w) × 2 m (h). It is highly recommended to have a one-piece poster on the board. For easier viewing, the poster length should not exceed 1.2m. Moreover, make sure you leave enough margins around the poster, i.e., do not cover the full board exactly. Discuss your work with the audience during your Poster Session. Poster Content The poster should include title, author name(s), affiliation(s), and an abstract. It is highly recommended to use illustrations, charts, tables, and other visually appealing objects to convey your research work. Make sure all materials are clearly readable from a distance. You may bring other relevant materials, e.g., full paper, addendum, etc. to your poster session. Poster Location Each board is labeled with a Poster ID. Please locate your Poster ID for your presentation. Each poster's presenter is responsible for setting up and taking down his/her own poster during the conference. Poster Set-Up and Take-Down You are requested to put up your poster before the morning coffee break and take it down no later than 15 minutes after the last session of the day. The organizers reserve the right to remove the poster by the appointed time. Posters must be attached with non-permanent adhesive (such as blue-tac or double-sided tape). Push-pins, thumb-tacks, or staples are not allowed. The organizers will provide such adhesive at the conference.
38
Roo
m40
240
340
440
540
640
7
8:30
- 17
:00
9:00
- 10
:30
Tuto
rial
10:3
0 -
10:4
5C
off
ee B
reak
10:4
5 -
12:1
5Tu
tori
al
12:1
5 -
14:0
0
14:0
0 -
15:3
0Tu
tori
al15
:30
- 15
:45
Co
ffee
Bre
ak15
:45
- 17
:15
Tuto
rial
18:0
0 -
20:0
0
Tues
day,
Oct
ober
3, 2
006
Reg
istr
atio
n
Xue
gon
g Z
HA
NG
Ale
x S
MO
LA
Am
ir A
TIY
A
Lun
ch B
reak
Ko
ji TS
UD
AR
usse
ll E
BE
RH
AR
TM
ahes
an N
IRA
NJA
N
Ope
ning
Cer
emon
y an
d W
elco
me
Rec
epti
onR
oom
401
Sess
ion
Sum
mar
y
39
Roo
m40
140
240
340
440
540
640
740
8
8:30
- 1
7:00
9:00
- 1
0:00
10:0
0 - 1
0:30
Invi
ted
Ses
sion
Ses
sion
Ses
sion
Ses
sion
Ses
sion
Ses
sion
Ses
sion
Ses
sion
WA
401
WA
402
WA
403
WA
404
WA
405
WA
406
WA
407
WA
408
AP
NN
A
Pre
side
ntia
lIn
vite
d Ta
lks
Neu
robi
olo
gy
Mod
elin
g an
d A
naly
sis
I
Lea
rnin
g A
lgo
rith
ms
I
Pat
tern
C
lass
ifica
tion
&
Rec
ogni
tion
I
PS
O I-
The
ory
&
Opt
imiz
atio
nF
inan
cial
A
pplic
atio
nsE
volu
tion
ary
Alg
orit
hms
I
Theo
reti
cal
Mod
elin
g a
nd
An
alys
is I
Nik
KA
SA
BO
VA
uckl
and
Un
iver
sity
of
Tech
nol
ogy
Tosh
io
AK
IMIT
SU
Uni
vers
ity
of
Toky
o
Rh
ee M
an K
ILK
AIS
T
Su
ngzo
on C
HO
Seo
ul N
atio
nal
Uni
vers
ity
Min
-jae
KA
NG
Che
ju N
atio
nal
U
nvie
rsity
Dal
ton
LUN
GA
Un
iver
sity
of
Witw
ater
sran
d
Kaz
ush
i IK
ED
AK
yoto
U
nive
rsit
y
Sad
ayu
ki
MU
RA
SH
IMA
Kag
osh
ima
Uni
vers
ity
DeL
iang
WA
NG
4496
4689
5190
4323
5452
4950
4208
Wlo
dzis
law
5015
4483
5320
4718
4597
4929
5350
DU
CH
5041
5170
4694
5321
4140
5077
5389
5284
4780
5439
4858
4838
5196
5377
4277
4583
5419
5210
4533
4545
5373
4945
4815
12:1
0 - 1
3:30
Wed
nesd
ay, O
ctob
er 4
, 200
6
Reg
istr
atio
n
Ple
nary
Sp
eech
Shu
n-ic
hi A
MA
RI
Roo
m 4
01C
hair
: DeL
iang
WA
NG
Cof
fee
Bre
ak
10:3
0 - 1
2:10
Par
alle
lS
essi
ons
Lun
ch B
reak
40
Invi
ted
IIS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onW
B40
1W
B40
2W
B40
3W
B40
4W
B40
5W
B40
6W
B40
7W
B40
8
Wo
rksh
opC
ogn
itive
P
roce
ssin
g /
Lear
ning
IS
VM
IP
atte
rn
Cla
ssifi
catio
n &
R
ecog
niti
on
II
Bio
med
ical
A
pplic
atio
ns I
Fore
cast
ing
&
Pre
dict
ion
IFu
zzy
Sys
tem
s I
AP
NN
A
Gov
erni
ng
Boa
rd M
eetin
g
Man
wai
MA
KH
ong
Ko
ng
Pol
ytec
hnic
U
nive
rsity
Min
ho L
EE
Kyu
ngp
ook
Nat
iona
l U
niv
ersi
ty
Ha-
Nam
N
GU
YE
NH
anku
k A
viat
ion
U
nive
rsity
Thom
as
VIL
LM
AN
NU
nive
rsity
Le
ipzi
g
Rin
ku
PA
NC
HA
LC
entr
al
Que
ensl
and
Uni
vers
ity
Mic
hael
SM
ALL
Hon
g K
ong
P
olyt
ech
nic
Uni
vers
ity
Fran
k LE
UN
GH
ong
Kon
g P
oly
tech
nic
Un
iver
sity
5580
5046
4874
5132
5200
5433
5532
4779
4567
5406
4574
5124
5511
5494
5233
4607
4835
5402
5042
4781
4478
4369
4558
4888
4814
5331
4968
5181
5116
15:1
0 - 1
6:30
13:3
0 - 1
5:10
Par
alle
lS
essi
ons
Cof
fee
Bre
ak
Wed
nesd
ay, O
ctob
er 4
, 200
6
41
Wed
nesd
ay, O
ctob
er 4
, 200
6
15:1
0 - 1
6:30
Invi
ted
IIIS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onW
C40
1W
C40
2W
C40
3W
C40
4W
C40
5W
C40
6W
C40
7W
C40
8
Wo
rksh
opC
ogn
itive
P
roce
ssin
g /
Lea
rnin
g II
Sp
ecia
l S
essi
on
New
Tre
nds
in
Sel
f-O
rgan
izin
g M
aps
I
Imag
e P
roce
ssin
g I
Spe
cial
S
essi
on
Com
puta
tiona
l In
telli
gen
ce in
E
cono
mic
s an
d
Fin
ance
Rob
otic
s an
d C
ontr
ol I
Dat
a P
repr
oce
ssin
g-Fe
atur
e S
elec
tion
I
AP
NN
A
Gov
erni
ng
Boa
rd M
eetin
g
Man
wai
MA
KH
ong
Ko
ng
Pol
ytec
hnic
U
nive
rsity
Van
essa
Y
AR
EM
CH
UK
Uni
vers
ity o
f A
lber
ta
Tets
uo
FUR
UK
AW
AK
yush
u In
stitu
te o
f Te
chno
logy
Ben
edic
te
BA
SC
LE,
Fra
nce
Tele
com
Shu
-Hen
g C
HE
NN
atio
nal
Ch
engc
hi
Uni
vers
ity
Nic
ola
s N
AV
ET
INR
IA
Wai
-keu
ng
FUN
GU
niv
ersi
ty o
f M
anito
ba
Tak
ash
i M
INO
AH
RA
Tak
usho
ku
Un
iver
sity
5287
5587
5423
5005
4823
5270
5240
5586
5093
5520
4896
5415
4771
5588
4897
5559
5296
4931
5542
5582
5398
5570
4266
5063
4457
5050
5579
4620
Pos
ter
Ses
sion
I
16:3
0 - 1
8:10
Par
alle
lS
essi
ons
42
Roo
m40
140
240
340
440
540
640
740
8
8:30
- 1
7:00
9:00
- 1
0:00
10:0
0 - 1
0:30
Invi
ted
IS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onT
A40
1T
A40
2T
A40
3T
A40
4T
A40
5T
A40
6T
A40
7T
A40
8
AP
NN
A
Pre
side
ntia
lIn
vite
d Ta
lks
Spe
cial
S
essi
on
Neu
rod
ynam
ics
SV
M II
Imag
e P
roce
ssin
g II
Opt
imiz
atio
nR
obo
tics
and
Con
trol
IIE
volu
tiona
ry
Alg
orith
ms
IIIC
A/B
SS
I
Laiw
an C
HA
NTh
e C
hin
ese
Uni
vers
ity o
f H
ong
Kon
g
Rub
in W
AN
GE
ast C
hin
a U
nive
rsity
of
Sci
ence
and
Te
chno
logy
Jam
es K
WO
KH
ong
Ko
ng
Un
iver
sity
of
Sci
ence
and
Te
chno
logy
Jaga
th
RA
JAP
AK
SE
Nan
yang
Te
chno
logi
cal
Un
iver
sity
Xia
olin
HU
The
Ch
ines
e U
nive
rsity
of
Hon
g K
ong
Jih-
Gau
JU
AN
GN
atio
nal
Ta
iwan
Oce
an
Un
iver
sity
Myu
ng W
on
KIM
So
ong
sil
Uni
vers
ity
Kun
ZH
AN
GTh
e C
hine
se
Un
iver
sity
of
Hon
g K
ong
Shi
ro U
sui
6001
5471
4572
5025
5249
5261
5521
Sho
ujue
Wan
g55
5450
0454
7050
5542
9147
0155
6348
4052
2348
2551
1041
6352
0452
4355
0252
8651
3551
9748
0844
5949
5350
4450
6248
2844
6953
26
12:1
0 - 1
3:30
Thur
sday
, Oct
ober
5, 2
006
Reg
istr
atio
n
Ple
nary
Sp
eech
Sw
arm
Inte
llig
ence
an
d E
xten
ded
An
alog
Co
mp
utin
gR
uss
ell E
BE
RH
AR
T
Roo
m 4
01C
hair
: Ir
win
KIN
G
Cof
fee
Bre
ak
10:3
0 - 1
2:10
Par
alle
lS
essi
ons
Lun
ch B
reak
43
Thur
sday
, Oct
ober
5, 2
006
Invi
ted
IIS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onT
B40
1T
B40
2T
B40
3T
B40
4T
B40
5T
B40
6T
B40
7T
B40
8
AP
NN
A
Pre
side
ntia
lIn
vite
d Ta
lks
Cog
niti
ve
Pro
cess
ing
/ Le
arn
ing
III
Spe
cial
S
essi
on
Ext
ensi
ons
of
Inde
pen
dent
C
ompo
nen
t A
nal
ysis
I
Pat
tern
C
lass
ifica
tion
&
Rec
ogni
tion
III
(Fac
e A
naly
sis
& P
roce
ssin
g)
Bio
med
ical
A
pplic
atio
ns II
Lear
nin
g A
lgo
rith
ms
IIFu
zzy
Sys
tem
s II
Theo
reti
cal
Mod
elin
g a
nd
An
alys
is II
Shi
ro U
SU
IR
IKE
N
Min
ho L
EE
Kyu
ngp
ook
Nat
iona
l U
niv
ersi
ty
Kun
ZH
AN
GTh
e C
hine
se
Un
iver
sity
of
Hon
g K
ong
Sal
im
BO
UZ
ER
DO
UM
Un
iver
sity
of
Wol
long
ong
Mic
hael
SM
ALL
Ho
ng K
ong
P
olyt
echn
ic
Uni
vers
ity
Sab
ri A
RIK
Ista
nbu
l U
niv
ersi
ty
Yi Z
HO
UN
anya
ng
Tech
nol
ogi
cal
Uni
vers
ity
Man
wai
MA
KH
ong
Ko
ng
Pol
ytec
hnic
U
nive
rsity
Soo
-You
ng L
ee53
4949
1546
3754
2643
9652
2755
80N
ik K
asab
ov55
1655
6150
1947
9343
9552
3253
0246
6955
4751
7353
1844
2352
5755
2454
4355
5149
0548
3049
1749
3754
6746
0955
5842
2550
51
15:1
0 - 1
6:30
13:3
0 - 1
5:10
Par
alle
lS
essi
ons
Cof
fee
Bre
ak
44
Thur
sday
, Oct
ober
5, 2
006
15:1
0 - 1
6:30
Invi
ted
IIIS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onT
C40
1T
C40
2T
C40
3T
C40
4T
C40
5T
C40
6T
C40
7T
C40
8
K.C
. Won
g Fo
und
atio
nIn
vite
d Ta
lks
Cog
niti
ve
Pro
cess
ing
/ L
earn
ing
IV
Har
dwar
e Im
plem
enta
tion
Pat
tern
C
lass
ifica
tion
&
Rec
ogn
itio
n IV
(F
ace
Ana
lysi
s &
Pro
cess
ing)
PS
O II
-A
ppl
icat
ion
sFo
reca
stin
g &
P
redi
ctio
n II
Dat
a P
repr
ocss
ing-
Dim
ensi
on
Red
uctio
n II
Theo
reti
cal
Mod
elin
g a
nd
Ana
lysi
s III
Jun
WA
NG
The
Chi
nese
U
niv
ersi
ty o
f H
ong
Ko
ng
Qi Z
HA
NG
Son
y C
ompu
ter
Sci
ence
La
bor
ator
ies
Tosh
i SA
ITO
Ho
sei
Uni
vers
ity
Chr
isto
pher
Y
AN
GTh
e C
hine
se
Uni
vers
ity o
f H
ong
Kon
g
San
ghou
n O
HG
wan
gju
In
stitu
te o
f S
cien
ce a
nd
Tec
hno
logy
Xu
n L
IAN
GP
ekin
g U
nive
rsity
Zen
glin
XU
The
Chi
nes
e U
nive
rsity
of
Ho
ng K
ong
Tin
gw
en
HU
AN
GTe
xas
A&
M
Uni
vers
ity a
t Q
atar
Aik
e G
uo
4735
5435
5013
5504
4472
5264
4969
Tian
you
Ch
ai49
7152
1854
2142
8251
3752
9548
8646
5749
5648
7752
8850
4753
7648
1744
2254
4050
7542
8449
8651
8353
6750
8747
20
18:3
0 - 1
9:00
19:0
0 - 2
1:00
Pos
ter
Ses
sion
II
16:3
0 - 1
8:10
Par
alle
lS
essi
ons
Rec
eptio
n
Ban
que
t
45
Roo
m40
140
240
340
440
540
640
740
8
8:30
- 1
7:00
9:00
- 1
0:00
10:0
0 - 1
0:30
Frid
ay, O
ctob
er 6
, 200
6
Reg
istr
atio
n
PA
NE
L S
ES
SIO
N
The
Futu
re o
f N
eura
l Inf
orm
atio
n P
roce
ssin
gLa
iwan
CH
AN
, Wlo
dzis
law
DU
CH
, YiX
in Z
HO
NG
, and
Nik
KA
SA
BO
V
Roo
m 4
01M
ode
rato
r: Ir
win
KIN
G
Cof
fee
Bre
ak
46
Invi
ted
IS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onFA
401
FA40
2F
A40
3FA
404
FA
405
FA40
6FA
407
FA
408
AP
NN
A
Pre
side
ntia
lIn
vite
d Ta
lks
Spe
cial
S
essi
on
New
Tre
nds
in
Sel
f-O
rgan
izin
g
Map
s II
Sig
nal
Pro
cess
ing
Spe
cial
S
essi
on
Incr
emen
tal
Sel
f-O
rgan
izin
g N
etw
orks
Spe
cial
S
essi
on
Ad
vanc
es in
M
ach
ine
Lea
rnin
g M
eth
ods
base
d P
atte
rn
Rec
ogni
tion
Dat
a &
Tex
t P
roce
ssin
gC
om
put
er V
isio
nTh
eore
tica
l M
odel
ing
and
A
nal
ysis
IV
Man
wai
MA
KH
ong
Ko
ng
Pol
ytec
hnic
U
nive
rsity
Tets
uo
FU
RU
KA
WA
Kyu
shu
Inst
itute
of
Tech
nolo
gy
Sh
ouyu
an
YA
NG
Jian
gxi
Un
iver
sity
of
Fina
nce
&
Eco
nom
ics
Sam
an
HA
LGA
MU
GE
,A
rthu
r H
SU
Un
iver
sity
of
Mel
bou
rne
Dam
min
da
AL
AH
AK
OO
NM
onas
h U
nive
rsity
Kai
zhu
HU
AN
GF
ujits
u C
hin
a
Man
uel
MA
RT
IN-
ME
RIN
OU
niv
ersi
dad
P
ontif
icia
de
Sal
aman
ca
Usm
an K
HA
NB
rune
l U
niv
ersi
ty
And
rew
LE
UN
GC
ity U
nive
rsity
o
f Hon
g K
ong
Yix
in Z
hon
g55
9048
5955
3855
4653
9754
5045
84La
iwan
Cha
n55
8351
2055
3755
6045
8847
8844
8555
8155
6845
4655
6255
4350
4954
6250
7451
7852
7555
4052
6354
4553
3252
5648
2644
9052
3951
8953
4644
2746
9154
12
12:1
0 - 1
3:30
10:3
0 - 1
2:10
Par
alle
lS
essi
ons
Lun
ch B
reak
Frid
ay, O
ctob
er 6
, 200
6
47
Invi
ted
IIS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
onS
essi
on
FB40
1FB
402
FB
403
FB40
4F
B40
5FB
406
FB40
7F
B40
8
Invi
ted
Talk
s
Dat
a P
roce
ssin
g &
In
form
atio
n S
ecur
ity
Pat
tern
C
lass
ifica
tion
&
Rec
ogni
tion
V
Bio
med
ical
A
ppl
icat
ion
s III
Dat
a &
Tex
t P
roce
ssin
gM
anu
fact
urin
g A
pplic
atio
ns
Theo
reti
cal
Mod
elin
g a
nd
Ana
lysi
s V
Laiw
an C
HA
NTh
e C
hin
ese
Uni
vers
ity o
f H
ong
Kon
g
Taka
yuki
K
IMU
RA
Sai
tam
a U
nive
rsity
Eri
c Y
UA
STR
I
Cha
ojia
n S
HI
Sha
ngh
ai
Mar
itim
e U
nive
rsity
Che
ngh
ua L
IC
hon
buk
Nat
ion
al
Un
iver
sity
Sh
uich
i K
UR
OG
IK
yusu
Inst
itute
of
Tec
hno
log
y
Hai
xuan
YA
NG
The
Ch
ines
e U
nive
rsity
of
Hon
g K
ong
Aap
o49
8143
0154
8751
4045
4451
03 H
YV
AR
INE
N48
9348
9251
7946
2748
1643
9752
5550
0250
9155
3652
0546
9053
8555
7453
19
15:0
0 - 1
6:30
13:3
0 - 1
5:10
Par
alle
lS
essi
ons
Pos
ter
Ses
sion
III a
nd
Far
ewel
l Par
ty
Frid
ay, O
ctob
er 6
, 200
6
WA401
48
Wednesday Session A, 4 October 2006 WA401 Invited talk by DeLiang Wang Invited talk by Wlodzislaw Duch Section Chair: Nik KASABOV, Auckland University of Technology
WA402
49
WA402 Neurobiology Modeling and Analysis I Section Chair: Toshio AKIMITSU University of Tokyo WA402-1 PID: 4496 How reward can induce reverse replay of behavioral sequences in the hippocampus Colin Molter, RIKEN Brain Science Institute, Japan Naoyuki Sato, RIKEN Brain Science Institute, Japan Utku Salihoglu, Brussels University, Belgium Yoko Yamaguchi, RIKEN Brain Science Institute, Japan In a recent experiment, Foster and Wilson [1] have observed reverse replay of behavioral sequences in rodents’ hippocampal place cells during non-running awake state in coincidence with sharp waves. In this paper, to elucidate this reverse replay mechanism, a theta phase precession computational model is assumed in one time trial learning experiment of a behavioral sequence. Our simulations demonstrate that reverse replay can occur during sharp waves states under the assumption that place cells’ excitability is elevated by reward. This reward induced reverse replay in the hippocampus might serve as a basis for reinforcement learning. WA402-2 PID: 5015 BCM-type synaptic plasticity model using a linear summation of calcium elevations as a sliding threshold Hiroki Kurashige, Tamagawa University, Japan Yutaka Sakai, Tamagawa University, Japan It has been considered that an amount of calcium elevation in a synaptic spine determines whether the synapse is potentiated or depressed. However, it has been pointed out that simple application of the principle can not reproduce properties of spike-timing-dependent plasticity (STDP). To solve the problem, we present a possible mechanism using dynamically sliding threshold as the linear summation of calcium elevations induced by single pre-synaptic and post-synaptic spikes. We demonstrate that the model can reproduce the timing dependence of biological STDP. In addition, we find that the model can reproduce the dependence of biological STDP on the initial synaptic strength, which is found to be asymmetric for synaptic potentiation and depression, whereas no explicit initial-strength dependence or asymmetric mechanism is incorporated into the model. WA402-3 PID: 5041 A New Method for Multiple Spike Train Analysis Based on Information Discrepancy Guang-Li Wang, Shanghai Jiao Tong University, China Xue Liu, Shanghai Jiao Tong University, China Pu-Ming Zhang, Shanghai Jiao Tong University, China Pei-Ji Liang, Shanghai Jiao Tong University, China Simultaneous recording of multiple spike trains from population of neurons provides the possibility for understanding how neurons work together in response to various
WA402
50
stimulations. But currently method is still lacking for researchers to perform multiple spike train data analysis and those existing techniques either allow people to analyze pairwise neuronal activities only or are seriously subject to the selection of parameters. In this paper, a new measurement of information discrepancy, which is based on the comparisons of subsequence distributions, is applied to deal with a group of spike trains (n > 2) and analyze the synchronization pattern among the neurons, where the analytical result mostly depends on the experimental data and is affected little by subjective interference. WA402-4 PID: 5284 Self-Organizing Network through Spike-Timing Dependent Plasticity Toshio Akimitsu,The University of Tokyo, Japan Akira Hirose, The University of Tokyo, Japan Yoichi Okabe, The University of the Air, Japan Many experimental results suggest that more precise spike timing is significant in neural information processing. From this point of view, we construct a self-organization model using the spatiotemporal patterns, where Spike-Timing Dependent Plasticity (STDP) tunes the conduction delays between neurons. STDP forms more smoothed map with the spatially random and dispersed patterns, whereas it causes spatially distributed clustering patterns from spatially continuous and synchronous inputs. These results suggest that STDP forms highly synchronous cell assemblies changing through external stimuli to solve a binding problem. WA402-5 PID: 4277 Ratio of average inhibitory to excitatory conductance modulates the response of Simple cell Akhil R Garg, J.N.V. University Jodhpur INDIA Basabi Bhaumik, I.I.T Delhi, INDIA Recent experimental study reports existence of complex type of interneurons in the primary visual cortex. The response of these inhibitory cells depends mainly upon feed-forward LGN inputs. The goal of this study is to determine the role of these cells in modulating the response of simple cells. Here we demonstrate that if the inhibitory contribution due to these cells balances the feed-forward excitatory inputs the spike response of cortical cell becomes sharply tuned. Using a single cell integrated and fire neuron model we show that the ratio of average inhibitory to excitatory conductance controls the balance between excitation and inhibition. We find that many different values of ratio can result in balanced condition. However, the response of the cell is not sharply tuned for each of these ratios. In this study we explicitly determine the best value of ratio needed to make the response of the cell sharply tuned. WA402-6 PID: 4945 Comparison of spike-train responses of a pair of coupled neurons under the external stimulus Wuyin Jin, Lanzhou University of Technology, China Yaobing Wei, Lanzhou University of Technology, China
WA402
51
Zhiyuan Rui, Lanzhou University of Technology, China Changfeng Yan, Lanzhou University of Technology, China Numerical calculations have been made on the consistent spike-train response of a pair of locus ceruleus (LC) neurons coupled by synapse. The coupled, excitable LC neurons are assumed to receive the constant, periodic and chaotic external stimulus at dendrite of the neuron, and whose soma potential being adopted to driving the other one along axon. With appropriated stimulus and coupling strength, the synchronization oscillation between the two neurons is well preserved even when the external stimulus is chaotic, for the small time scale stimulus, one inspiring simulations results, the wave shape or chaotic attractor of stimulus could be transmitted completely by neuronal ISIs sequence, including the periodic, chaotic characters of stimulus, such as, phase space or chaotic attractors, but this phenomenon disappears for big time scale stimulus.
WA403
52
WA403 Learning Algorithms I Section Chair: Rhee Man KIL, KAIST WA403-1 PID: 4689 Adaptive Kernel Leaning Networks for Nonlinear Process Identification Haiqing Wang, Zhejiang University, China Ping Li, Zhejiang University, China Zhihuan Song, Zhejiang University, China Steven X. Ding, University of Duisburg-Essen, Germany By kernelizing the traditional least-square based identification method, an adaptive kernel learning (AKL) network is proposed for nonlinear process modeling, which utilizes kernel mapping and geometric angle to build the network topology adaptively. The generalization ability of AKL network is controlled by introducing a regularized optimization function. Two forms of learning strategies are addressed and their corresponding recursive algorithms are derived. Numerical simulations show this simple AKL networks can learn the process nonlinearities with very small samples, and has excellent modeling performance in both the deterministic and stochastic environments. WA403-2 PID: 4483 Localized Bayes Estimation for Non-identifiable Models Shingo Takamatsu, Tokyo Institute of Technology, Japan Shinichi Nakajima, Nikon Corporation, Japan Sumio Watanabe, Tokyo Institute of Technology, Japan Hierarchical learning machines such as neural networks are now being used in many applications. Although the Bayes ensemble learning gives the good generalization performance in such hierarchical learning machines, it is difficult to realize the posterior distribution because of the singularities in the parameter space. In this paper, we propose a new learning algorithm which enables us to construct a localized posterior distribution. We call this method Localized Bayes estimation and theoretically show that it attains the smaller generalization error in reduced rank approximations. WA403-3 PID: 5170 Optimality of Kernel Density Estimation of Prior Distribution in Bayes Network Hengqing Tong, Wuhan University of Technology, China Yanfang Deng, Wuhan University of Technology, China Ziling Li, Wuhan University of Technology, China The key problem of inductive-learning in Bayes network is the estimator of prior distribution. This paper adopts general naive Bayes to handle continuous variables, and proposes a kind of kernel function constructed by orthogonal polynomials, which is used to estimate the density function of prior distribution in Bayes network. The paper then makes further researches into optimality of the kernel estimation of density and derivatives. When the sample is fixed, the estimators can keep continuity and smoothness,
WA403
53
and when the sample size tends to infinity, the estimators can keep good convergence rates. WA403-4 PID: 4780 Learning Hierarchical Bayesian Networks for Large-Scale Data Analysis Kyu-Baek Hwang, Soongsil University Byoung-Hee Kim, Seoul National University, Korea Byoung-Tak Zhang, Seoul National University, Korea Bayesian network learning is a useful tool for exploratory data analysis. However, applying Bayesian networks to the analysis of large-scale data, consisting of thousands of attributes, is not straight- forward because of the heavy computational burden in learning and visualization. In this paper, we propose a novel method for large-scale data analysis based on hierarchical compression of information and con- strained structural learning, i.e., hierarchical Bayesian networks (HBNs). The HBN can compactly visualize global probabilistic structure through a small number of hidden variables, approximately representing a large number of observed variables. An efficient learning algorithm for HBNs, which incrementally maximizes the lower bound of the likelihood function, is also suggested. The effectiveness of our method is demonstrated by the experiments on synthetic large-scale Bayesian networks and a real-life microarray dataset. WA403-5 PID: 4583 Confidence Intervals for the Risks of Regression Models Imhoi Koo, Korea Advanced Institute of Science and Technology, Korea; Rhee Man Kil, Korea Advanced Institute of Science and Technology, Korea The empirical risks of regression models are not accurate since they are evaluated from the finite number of samples. In this context, we investigate the confidence intervals for the risks of regression models, that is, the intervals between the expected and empirical risks. The suggested method of estimating confidence intervals can provide a tool for predicting the performance of regression models.
WA404
54
WA404 Pattern Classification & Recognition I Section Chair: Sungzoon CHO, Seoul National University WA404-1 PID: 5190 Performance enhancement of extreme learning machine for multi-category sparse data classification S. Suresh, Nanyang Technological University, Singapore S. Saraswathi, Nanyang Technological University, Singapore N. Sundararajan, Nanyang Technological University, Singapore This paper presents a performance enhancement schemes for the recently developed Extreme Learning Machine (ELM) for multi-category sparse data classification problems. ELM is a single hidden layer neural network with good generalization capabilities and extremely fast learning. In ELM network, input weights are randomly chosen and output weights are analytically calculated. The generalization performance of the ELM algorithm depends on the following free parameters: number of hidden neurons, input weights, and bias values. Selection of these parameters for the best performance of ELM is a complex problem. In this paper, we present a method to search for the best parameters of ELM using a K-fold validation scheme. For evaluating the performance of the proposed method, we consider a multi-class sparse human cancer detection problem using micro-array gene expression data. Performance of the proposed scheme is compared with an evolving ELM algorithm and also with the support vector machine results reported in the literature. The results clearly show the superior performance of the proposed approach. WA404-2 PID: 5320 Using One-Class Classifier Ensembles for Multi-Class Classification Tao Ban, Kobe University, Japan Shigeo Abe, Kobe University, Japan In this paper we address the problem of how to implement a multiclass classifier by an ensemble of one-class classifiers. One-class classifiers are first trained for each class and then a decision function is formulated based on minimum distance rules. Two kinds of one-class classifiers are explored: the Support Vector Domain Description and a Kernel Principle Component Analysis based method. Both of the methods can work in the feature space and deal with nonlinear classification problems. Experiments on some benchmark datasets show that the proposed methods with carefully tuned parameters have comparable generalization ability with Support Vector Machines while having some other advantages. WA404-3 PID: 4694 Distance Function Learning in Error-Correcting Output Coding Framework Dijun Luo, Zhejiang University Rong Xiong, Zhejiang University
WA404
55
This paper presents a novel framework of error-correcting output coding (ECOC) addressing the problem of multi-class classification. By weighting the output space of each base classifier which is trained independently, the distance function of decoding is adapted so that the samples are more discriminative. A criterion generated over the Extended Pair Samples (EPS) is proposed to train the weights of output space. Some properties still hold in the new framework: any classifier, as well as distance function, is still applicable. We first conduct empirical studies on UCI datasets to verify the presented framework with four frequently used coding matrixes and then apply it in RoboCup domain to enhance the performance of agent control. Experimental results show that our supervised learned decoding scheme improves the accuracy of classification significantly and betters the ball control of agents in a soccer game after learning from experience. WA404-4 PID: 5439 Combining pairwise coupling classifiers using individual logistic regressions NobuhikoYamaguchi, Saga University, Japan Pairwise coupling is a popular multi-class classification approach that prepares binary classifiers separating each pair of classes, and then combines the binary classifiers together. This paper proposes a pairwise coupling combination strategy using individual logistic regressions (ILR-PWC). We show analytically and experimentally that the ILR-PWC approach is more accurate than the individual logistic regressions. WA404-5 PID: 5419 The Novelty Detection Approach for Different Degrees of Class Imbalance Hyoung-joo Lee, Seoul National University, Korea Sungzoon Cho, Seoul National University, Korea We show that the novelty detection approach is a viable solution to the class imbalance and examine which approach is suitable for different degrees of imbalance. In experiments using SVM-based classifiers, when the imbalance is extreme, novelty detectors are more accurate than balanced and unbalanced binary classifiers. However, with a relatively moderate imbalance, balanced binary classifiers should be employed. In addition, novelty detectors are more effective when the classes have a non-symmetrical class relationship.
WA405
56
WA405 PSO I (Theory & Optimization) Section Chair: Min-jae KANG, Cheju National University WA405-1 PID: 4323 Quantum-Behaved Particle Swarm Optimization for Integer Programming Jing Liu, Southern Yangtze University, China Jun Sun, Southern Yangtze University, China Wenbo Xu, Southern Yangtze University, China Based on our previously proposed Quantum-behaved Particle Swarm Optimization (QPSO), this paper discusses the applicability of QPSO to integer programming. QPSO is a global convergent search method, while the original Particle Swarm (PSO) cannot be guaranteed to find out the optima solution of the problem at hand. The application of QPSO to integer programming is the first attempt of the new algorithm to discrete optimization problem. After introduction of PSO and detailed description of QPSO, we propose a method of using QPSO to solve integer programming. Some benchmark problems are employed to test QPSO as well as PSO for performance comparison. The experiment results show the superiority of QPSO to PSO on the problems. WA405-2 PID: 4718 Neural Network Training Using Stochastic PSO Xin Chen, University of Macau Yangmin Li, University of Macau Particle swarm optimization is widely applied for training neural network. Since in many applications the number of weights of NN is huge, when PSO algorithms are applied for NN training, the dimension of search space is so large that PSOs always converge prematurely. In this paper an improved stochastic PSO (SPSO) is presented, to which a random velocity is added to improve particles’ exploration ability. Since SPSO explores much thoroughly to collect information of solution space, it is able to find the global best solution with high opportunity. Hence SPSO is suitable for optimization about high dimension problems, especially for NN training. WA405-3 PID: 5321 Solving Multiprocessor Real-Time System Scheduling with Enhanced Competitive Scheme Ruey-Maw Chen, National Chin-yi Institute of Technology, Taiwan Shih-Tang Lo,National Cheng-Kung University, Taiwan Yueh-Min Huang, National Cheng-Kung University, Taiwan A new method based on Hopfield Neural Networks (HNN) for solving real-time scheduling problem is adopted in this study. Neural network using competitive learning rule provides a highly effective method and deriving a sound solution for scheduling problem. Moreover, competitive scheme reduces network complexity. However, competitive scheme is a 1-out-of-N confine rule and applicable for limited scheduling problems. Restated, the processor may not be full utilization for scheduling problems. To facilitate the non-fully utilized problem, extra neurons are introduced to the
WA405
57
Competitive Hopfield Neural Network (CHNN). Slack neurons are imposed on CHNN with respected to pseudo processes. Simulation results reveal that the competitive neural network imposed on the proposed energy function with slack neurons integrated ensures an appropriate approach of solving both full and non-full utilization multiprocessor real-time system scheduling problems. WA405-4 PID: 4858 A Distributed Hybrid Algorithm for Optimized Resource Allocation Problem Kyeongmo Park, The Catholic University Sungcheol Kim, Sangmyung University Chuleui Hong, Sangmyung University This paper presents a novel distributed Mean-field Genetic algorithm called MGA for the load balancing problems in MPI environments. The proposed MGA is a hybrid algorithm of Mean Field Annealing (MFA) and Simulated-annealing-like Genetic Algorithm (SGA). The proposed MGA combines the benefit of rapid convergence property of MFA and the effective genetic operations of SGA. Our experimental results indicate that the composition of heuristic mapping methods improves the performance over the conventional ones in terms of communication cost, load imbalance and maximum execution time. It is also proved that the proposed distributed algorithm maintains the convergence properties of sequential algorithm while it achieves almost linear speedup as the problem size increases. WA405-5 PID: 5210 Neural Networks for Optimization Problem with Nonlinear Constraints Min-jae Kang, Cheju National University, Korea Ho-chan Kim, Cheju National University, Korea Farrukh Aslam Khan, Cheju National University, Korea Wang-cheol Song, Cheju National University, Korea Sangjoon Lee, Cheju National University, Korea Hopfield introduced the neural network for linear programming with linear constraints. In this paper, Hopfield neural network has been generalized to solve the optimization problems including nonlinear constraints. The proposed neural network can solve a nonlinear cost function with nonlinear constraints. Also, methods have been discussed to reconcile optimization problems with neural networks and implementation of the circuits. Simulation results show that the computational energy function converges to stable point by decreasing the cost function as the time passes.
WA406
58
WA406 Financial Applications Section Chair: Dalton LUNGA, University of Witwatersrand WA406-1 PID: 5452 NN-OPT: Neural Network for Option Pricing Using Multinomial Tree Hung-Ching (Justin) Chen, Rensselaer Polytechnic Institute, United States of America Malik Magdon-Ismail, Rensselaer Polytechnic Institute, United States of America We provide a framework for learning to price complex options by learning risk-neutral measures (Martingale measures). In a simple geometric Brownian motion model, the price volatility, fixed interest rate and a no-arbitrage condition suffice to determine a unique risk-neutral measure. On the other hand, in our framework, we relax some of these assumptions to obtain a class of allowable risk-neutral measures. We then propose a framework for learning the appropriate risk-neural measure. In particular, we provide an efficient algorithm for back-propagating gradients through multinomial pricing trees. Since the risk-neutral measure prices all options simultaneously, we can use all the option contracts on a particular stock for learning. We demonstrate the performance of these models on historical data. Finally, we illustrate the power of such a framework by developing a real time trading system based upon these pricing methods. WA406-2 PID: 4597 A Brain-Inspired Cerebellar Associative Memory Approach to Option Pricing and Arbitrage Trading S. D. Teddy, Nanyang Technological University, Singapore E. M-K. Lai, Nanyang Technological University, Singapore C. Quek, Nanyang Technological University, Singapore Option pricing is a process to obtain the theoretical fair value of an option based on the factors affecting its price. Currently, the nonparametric and computational methods of option valuation are able to construct a model of the pricing formula from historical data. However, these models are generally based on a global learning paradigm, which may not be able to efficiently and accurately capture the dynamics and time-varying characteristics of the option data. This paper proposes a novel brain-inspired cerebellar associative memory model for pricing American-style option on currency futures. The proposed model, called PSECMAC, constitutes a local learning model that is inspired by the neurophysiological aspects of the human cerebellum. The PSECMAC-based option pricing model is subsequently applied in a mis-priced option arbitrage trading system. Simulation results show a return on investment as high as 23.1% for a relatively risk-free investment. WA406-3 PID: 4140 A Reliability-based RBF Network Ensemble Model for Foreign Exchange Rates Predication Lean Yu, Chinese Academy of Sciences, China Wei Huang, Huazhong University of Science and Technology, China Kin Keung Lai, City University of Hong Kong, China
WA406
59
Shouyang Wang, Chinese Academy of Sciences, China In this study, a reliability-based RBF neural network ensemble forecasting model is proposed to overcome the shortcomings of the existing neural ensemble methods and ameliorate forecasting performance. In this model, the ensemble weights are determined by the reliability measure of RBF network output. For testing purposes, we compare the new ensemble model’s performance with some existing network ensemble approaches in terms of three exchange rates series. Experimental results reveal that the prediction using the proposed approach is consistently better than those obtained using the other methods presented in this study in terms of the same measurements. WA406-4 PID: 4838 Pricing Options in Hong Kong Market Based on Neural Networks Xun Liang, Peking University, China Haisheng Zhang, Peking University, China Jian Yang, Peking University, China Option pricing is one of the important issues in the financial industry and has been studied for decades. Many classical and successful pricing models have been presented to implement the pricing processing either by numerical computing or by simulation. In this paper, a new option pricing model based on a three-layer feedforward neural network is established to improve the pricing performance. The new model combines 4 traditional pricing models to obtain a better forecasting result based on learning and cutting down their forecasting errors. Numerical experiments are conducted on the data of Hong Kong option market from March 2005 to July 2005. The new model improves the pricing performance remarkably compared to the traditional option pricing models. WA406-5 PID: 4533 Neural networks, Fuzzy Inference Systems and Adaptive-Neuro Fuzzy Inference systems for financial decision making Pretesh B. Patel, University of the Witwatersrand, South Africa Tshilidzi Marwala, University of the Witwatersrand, South Africa. This paper employs pattern classification methods for assisting investors in making financial decisions. Specifically, the problem entails the categorization of investment recommendations. Based on the forecasted performance of certain indices, the Stock Quantity Selection Component is to recommend to the investor to purchase stocks, hold the current investment position or sell stocks in possession. Three designs of the component were implemented and compared in terms of their complexity as well as scalability. Designs that utilized 1, 4 and 16 classifiers, respectively, were developed. These designs were implemented using Artificial Neural Networks, Fuzzy Inference Systems as well as Adaptive Neuro-Fuzzy Inference Systems. The design that employed 4 classifiers achieved low complexity and high scalability. As a result, this design is most appropriate for the application of concern. WA406-6 PID: 4815
WA406
60
Online Forecasting of Stock Market Movement Direction Using the Improved Incremental Algorithm Dalton Lunga, University of the Witwatersrand Tshilidzi Marwala, University of the Witwatersrand In this paper we present a particular implementation of the Learn++ algorithm: we investigate the predictability of financial movement direction with Learn++ by forecasting the daily movement direction of the Dow Jones. The Learn++ algorithm is derived from the Adaboost algorithm, which is denominated by sub-sampling. The goal of concept learning, according to the probably approximately correct weak model, is to generate a description of another function, called the hypothesis, which is close to the concept, by using a set of examples. The hypothesis which is derived from weak learning is boosted to provide a better composite hypothesis in generalizing the establishment of the final classification boundary. The framework is implemented using multi-layer Perceptron (MLP) as a weak Learner. First, a weak learning algorithm, which tries to learn a class concept with a single input Perceptron, is established. The Learn++ algorithm is then applied to improve the weak MLP learning capacity and introduces the concept of online incremental learning. The proposed framework is able to adapt as new data are introduced and is able to classify.
WA407
61
WA407 Evolutionary Algorithms I Section Chair: Kazushi IKEDA, Kyoto University WA407-1 PID: 4950 Synthesis of Desired Binary Cellular Automata through the Genetic Algorithm Satoshi Suzuki, Hosei University, Japan Toshimichi Saito, Hosei University, Japan This paper presents a GA-based synthesis algorithm of a cellular automaton (CA) that can generate a desired spatio-temporal pattern. Time evolution of CA is determined by a rule table the number of which is enormous even for relatively small size CAs: the brute-force search is almost impossible. In our GA-based synthesis algorithm, a gene corresponds to a rule and a masking technique is used to preserve gene(s) with good fitness. Performing basic numerical experiments we have confirmed that the masking works effectively and the algorithm can generate a desired rule table. We have also considered an application to reduction of noise inserted randomly to a spatio-temporal pattern. WA407-2 PID: 4929 On Properties of Genetic Operators from a Network Analytical Viewpoint Hiroyuki Funaya, Kyoto University, Japan Kazushi Ikeda, Kyoto University, Japan In recent years, network analysis has revealed that some real networks have the properties of small-world and/or scale-free networks. In this paper, a simple Genetic Algorithm (GA) is regarded as a network where each node and each edge respectively represent a population and the possibility of the transition between two nodes. The characteristic path length, which is one of the most popular criteria in small-world networks, is derived analytically. The results show how the crossover operation works in GAs to shorten the path length between two populations, compared to the length of the network with the mutation operation. WA407-3 PID: 5077 SDMOGA: A New Multi-Objective Genetic Algorithm Based on Objective Space Divided Wangshu Yao, Soochow University, China Chen Shifu, Nanjing University, China Chen Zhaoqian, Nanjing University, China Most contemporary multi-objective evolutionary algorithms (MOEAs) have high computational demand. In this paper, a new MOEA based on objective space divided named SDMOGA is proposed. SDMOGA transforms the Pareto ranking into the sum of interval index ranking among individuals in objective space divided, and uses a method of individual crowding operator similar to adaptive grid to keep population diversity. Experimental results on four nicely balance functions show that SDMOGA has high efficiency, low run-time complexity and good convergence. WA407-4
WA407
62
PID: 5196 Hamming Sphere Solution Space Based Genetic Multi-user Detection Lili Lin, Zhejiang Gongshang University, China Many researches on genetic algorithm based multi-user detection indicate initial population has crucial effects on the performance of detectors. Commonly used method to obtain initial population is to perturb the input chromosome randomly, which fails to fully exploit the effective information delivered by input chromosome. This paper proposes a kind of Hamming Sphere Solution Space based Genetic Multi-User Detector (HSSSGMUD), which constructs the initial population in a simple but effective manner. Firstly, select the input chromosome, and regard it as the center of a sphere in PK dimensions space, where P is data packet length and K is user numbers in Code Division Multiple Access (CDMA) system. Then, the concept of Hamming sphere space is used to obtain other chromosomes of initial population. Simulation results show the proposed HSSSGMUD not only achieves lower Bit Error Ratio (BER) and better near-far resistant ability, but also converges quickly. WA407-5 PID: 4545 Decision Tree Classifier Based-Genetic Algorithm for Remote Sensing Mapping with SPOT-5 Data in HongShiMao Watershed of Loess Plateau Region, China Mingxiang Huang, Chinese Academy of Sciences, China Jianhua Gong, Chinese Academy of Sciences, China Zhou Shi,Zhejiang University, China Lihui Zhang, Chinese Academy of Sciences, China The loess plateau is faced with severe soil erosion and run-off. Check-dams are the effective measures for soil and water conservation, concomitantly their planning and construction urgently require the current land use map. Remote sensing techniques play a key role for achieving up-to-date land use map. However, limited by the impact of hilly and gully terrain in the loess plateau, the commonly used classifier for remote sensing data can’t achieve satisfied results. In the paper, HongShiMao watershed in the loess plateau as the study area, Decision Tree Classifier(DTC) based on Genetic Algorithm(GA) was applied to the land use classification automatically. Compared with the results by Iterative Self-Organizing Data Analysis Technique (ISODATA), DTC based-GA had the far better results and its total accuracy was up to 83.2%. The results also show that bare field including barren, some grass and crop field which attribute to the soil erosion and runoff, covers the most part of the study area.
WA408
63
WA408 Theoretical Modeling and Analysis I Section Chair: Sadayuki MURASHIMA, Kagoshima University WA408-1 PID: 4208 New Results for Global Stability of Cohen-Grossberg Neural Networks with Discrete Time Delays Zeynep Orman, Istanbul University, Turkey Sabri Arik, Istanbul University, Turkey This paper studies the global convergence properties of Cohen-Grossberg neural networks with discrete time delays. Without assuming the symmetry of interconnection weight coefficients, and the monotonicity and dierentiability of activation functions, and by employing Lyapunov functions, we derive new delay independent sufficient conditions under which a delayed Cohen-Grossberg neural network converges to a globally asymptotically stable equilibrium point. Some examples are given to illustrate the advantages of the results over the previously reported results in the literature. WA408-2 PID: 5350 Exponential Stability of Neural Networks with Distributed Time Delays and Strongly Nonlinear Activation Functions Chaojin Fu, Hubei Normal University, China Zhongsheng Wang, ZhongYuan Institute of Technology, China In this paper, we provided a new technique based on the concept of comparison. Different from the Lyapunov method, the new technique showed that if the given conditions hold then the any state of neural networks with distributed time delays and strongly nonlinear activation functions is always bounded by exponential convergence function. In addition, some sufficient conditions are obtained to guarantee that such neural network is globally exponentially stable, or locally exponentially stable. Furthermore, we obtained the estimates of the exponential convergence rates and the region of exponential convergence. WA408-3 PID: 5389 Global stability of bidirectional associative memory neural networks with variable coefficients and S-type distributed delays Yonggui Kao, Ocean University of China, China Cunchen Gao, Ocean University of China, China Lu Wu, Ocean University of China, China Qinghe Ming, ZaoZhuang University, China This paper is devoted to investigation of the global asymptotic stability for Bidirectional associative memory (BAM) neural networks with variable coefficient and S-type distributed signal transmission delays along the axon of a neuron. Some sufficient conditions for global asymptotic stability of the networks were obtained, in which the boundedness and differentiability of the signal functions in some papers are deleted. Some examples are also presented to show that our results are new and improve the previous results.
WA408
64
WA408-4 PID: 5377 Convergence Study of Discrete Neural Networks with Delay Run-Nian Ma,Xi’an Jiaotong University, Dalian University, China Guo-Qiang Bai, Tsinghua University, China The convergence of discrete Hopfield neural networks with delay is mainly investigated and some results on the convergence are given. Several new sufficient conditions for the networks with delay converging towards a limit cycle with length at most 2 are obtained. Also, several conditions for the networks with delay converging towards a limit cycle with length 2 are investigated. All results established here extend the previous results on the convergence of both discrete Hopfield neural networks without delay and with delay in parallel updating mode. WA408-5 PID: 5373 The Perfect Recalling Rate of Morphological Associative Memory Ke Zhang,Kagoshima University, Japan Sadayuki Murashima, Kagoshima University, Japan Morphological associative memory is made up by replacing the sum of input signals by maximum or minimum operation. Ritter presented a procedure of associative memory, which utilizes associative memory matrix M and W sophisticatedly. This paper treats the recalling rate of associative memory. At first, conditions of perfect recalling are given. Then the perfect recalling rate of associative memory is derived based on these conditions. The formula of perfect recalling rate found out to be equal to the perfect recalling rate obtained by experiments.
WB401
65
Wednesday Session B, 4 October 2006
WB401 Workshop Section Chair: Manwai MAK, Hong Kong Polytechnic University WB401-1 PID: 5580 A solution to the Curse of Dimensionality Problem in Pairwise Scoring Techniques Man-Wai Mak, The Hong Kong Polytechnic University, Hong Kong Sun-Yuan Kung, Princeton University, United States of America This paper provides a solution to the curse of dimensionality problem in the pairwise scoring techniques that are commonly used in bioinformatics and biometrics applications. It has been recently discovered that stacking the pairwise comparison scores between an unknown patterns and a set of known patterns can result in feature vectors with nice discriminative properties for classification. However, such technique can lead to curse of dimensionality because the vectors size is equal to the training set size. To overcome this problem, this paper shows that the pairwise score matrices possess a symmetric and diagonally dominant property that allows us to select the most relevant features independently by an FDA-like technique. Then, the paper demonstrates the capability of the technique via a protein sequence classification problem. It was found that 10-fold reduction in the number of feature dimensions and recognition time can be achieved with just 4% reduction in recognition accuracy.
WB402
66
WB402 Cognitive Processing / Learning I Section Chair: Minho LEE, Kyungpook National University WB402-1 PID: 5046 Top-Down Attention Guided Object Detection Mei Tian, Beijing Jiaotong University, China Si-Wei Luo, Beijing Jiaotong University, China Ling-Zhi Liao, Beijing Jiaotong University, China Lian-Wei Zhao, Beijing Jiaotong University, China Existing attention models concentrate on bottom-up attention guidance, and lack of effective definition of top-down attention information. In this paper we define a new holistic scene representation and use it as top-down attention information which works in three ways. The first is to discriminate between close-up and open scene categories. The second and the third are to provide reliable priors for the presence or absence of object and the location of it. Compared with traditional attention guidance algorithms, our algorithm shows how scene classification and basing directly on entire scene without segmentation stages, facilitate the object detection. Two stages of pre-attention and focus attention enhance the detecting performance and are more suitable for vision information processing in high level. Experiment results prove the effectiveness of our algorithm. WB402-2 PID: 4779 Absolute quantification of brain Creatine concentration using long echo time PRESS sequence with an external standard and LCModel: verification with in vitro HPLC method Y.Lin, Shantou University Medical College, China Y.P. Zhang, Shantou University Medical College, China R.H.Wu, Shantou University Medical College, China H.Li, Shantou University Medical College, China Z.W.Shen, Shantou University Medical College, China X.K.Chen, Shantou University Medical College, China K.Huang, Shantou University Medical College, China G. Guo, Shantou University Medical College, China To investigate the accuracy for absolute quantification of brain creatine (Cr) concentration using in vivo long echo time PRESS sequence performed with an external standard. Ten swine and an external standard phantom were investigated by 1.5T GE Signa scanner and the standard head coil. 1H-MRS data were acquired from the two VOI (2x2x2cm3) placed in swine brain and external standard solution by using PRESS sequence with TE = 135 mses, TR = 1500 msec, and 128 scan averages. In vivo Cr evaluation was made by LCModel. In vitro Cr concentration was analyzed by HPLC method. In the 1H-MRS group, the Cr concentration was 9.37-10.137 mmol/kg. In the HPLC group, the Cr concentration was 8.905-10.126 mmol/kg. Good agreement was obtained between in these two methods (P=0.491), which indicated that long echo time PRESS sequence with an external standard can accurately detect brain Cr concentration. The application of LCModel introduces more convenience for the MRS quantification.
WB402
67
WB402-3 PID: 5494 Learning in Neural Network - Unusual Effects of Artificial Dreams Ryszard Tadeusiewicz, AGH University of Science and Technology, Poland Andrzej Izworski, AGH University of Science and Technology, Poland Most researchers focused on particular result ignore intermediate stages of learning process of neural networks. The unstable and transitory phenomena, discovered in neural networks during the learning process, long time after the initial stage of learning, when the network knows nothing because of random values of all weights, and long time before final stage of learning process, when the network knows (almost) everything, can be very interesting, especially when we can associate with them some psychological interpretations. Some “immature” neurons exhibit behavior that can be interpreted as source of “artificial dreams”. Article presents examples of simple neural networks with capabilities which might explain the origins of dreams and myths. WB402-4 PID: 4781 Spatial Attention in Early Vision Alternate Direction-of-Figure Nobuhiko Wagatsuma, University of Tsukuba, Japan Ko Sakai, University of Tsukuba, Japan We propose a computational model consisting of mutually connected V1, V2, and PP modules, to realize the effect of attention to the determination of border-ownership (BO) that tells on which side of a contour owns the border. The V2 module determines BO from surrounding contrast extracted by the V1 module that could be affected by top-down spatial attention from the PP module. The simulation results show that the spatial attention modifies the direction of figure, and that the direction of figure is even flipped in ambiguous figures such as the Rubin’s vase, although the attention is applied only to enhance local contrast in V1. These results show that the activities of BO selective cells in V2 are modified significantly when spatial attention functions in early visual area, V1. WB402-5 PID: 5331 Biologically Motivated Incremental Object Perception Based on Selective Attention Woong-Jae Won,Kyungpook National University, Korea Jiyoung Yeo, Kyungpook National University, Korea Sang-Woo Ban, Dongguk University, Korea Minho Lee, Kyungpook National University, Korea In this paper, we propose a biologically motivated object selective attention and object perception system, which was implemented by integrating a specific object preferable attention model with an incremental object perception model. The object oriented attention model can selectively pay attention to the candidates of an object in natural scenes based on a bottom-up selective attention model in conjunction with a top-down biased attention mechanism for a specific object. A generative model based on an incremental Bayesian parameter estimation is considered in order to perceive arbitrary objects in the attended areas. Combining an object oriented attention model with general object perception model, the developed system can not only pay attention to a specific target object but also memory the characteristics of task non-specific objects by
WB402
68
incremental manner. Experimental results show that the developed system generates good performance in successfully focusing on the target objects as well as incrementally perceiving objects in natural scenes.
WB403
69
WB403 SVM I Section Chair: Ha-Nam NGUYEN, Hankuk Aviation University WB403-1 PID: 4874 A Heuristic Weight-Setting Algorithm for Robust Weighted Least Squares Support Vector Regression Wen Wen, South China University of Technology, China Zhifeng Hao, Southeast University, China Zhuangfeng Shao, South China University of Technology, China Xiaowei Yang, South China University of Technology, China Ming Chen, Southeast University, China Firstly, a heuristic algorithm for labeling the “outlierness” of samples is presented in this paper. Then based on it, a heuristic weight-setting algorithm for least squares support vector machine (LS-SVM) is proposed to obtain the robust estimations. In the proposed algorithm, the weights are set according to the changes of the observed value in the neighborhood of a sample’s input space. Numerical experiments show that the heuristic weight-setting algorithm is able to set appropriate weights on noisy data and hence effectively improves the robustness of LS-SVM. WB403-2 PID: 4567 Feature Selection using SVM Probabilistic Outputs Kai Quan Shen, National University of Singapore, Singapore Chong Jin Ong, National University of Singapore, Singapore Xiao Ping Li, National University of Singapore, Singapore Hui Zheng, National University of Singapore, Singapore Einar P V Wilder-Smith, National University of Singapore, Singapore A ranking criterion based on the posterior probability is proposed for feature selection on support vector machines (SVM). This criterion has the ad-vantage that it is directly related to the importance of the features. Four approximations are proposed for the evaluation of this criterion. The performances of these approximations, used in the recursive feature elimination (RFE) approach, are evaluated on various artificial and real-world problems. Three of the proposed approximations show good performances consistently, with one having a slight edge over the other two. Their performances compare favorably with feature selection methods in the literature. WB403-3 PID: 5233 Unified Kernel Function and its Training Method for SVM Ha-Nam Nguyen, Hankuk Aviation University, Korea Syng-Yup Ohn, Hankuk Aviation University, Korea This paper proposes a unified kernel function for support vector machine and its learning method with a fast convergence and a good classification performance. We defined the unified kernel function as the weighted sum of a set of different types of basis kernel functions such as neural, radial, and polynomial kernels, which are trained by a new
WB403
70
learning method based on genetic algorithm. The weights of basis kernel functions in the unified kernel are determined in learning phase and used as the parameters in the decision model in the classification phase. The unified kernel and the learning method were applied to obtain the optimal decision model for the classification of two public data sets for diagnosis of cancer diseases. The experiment showed fast convergence in learning phase and resulted in the optimal decision model with the better performance than other kernels. Therefore, the proposed kernel function has the greater flexibility in representing a problem space than other kernel functions. WB403-4 PID: 4478 Parameterized Semi-supervised Classification Based on Support Vector for Multi-Relational Data Ling Ping, Jilin University, China Wang Zhe, Jilin University, China Zhou Chunguang, Jilin University, China A Spectrum-based Support Vector Algorithm (SSVA) to resolve semi-supervised classification for relational data is presented in this paper. SSVA extracts data representatives and groups them with spectral analysis. Label assignment is done according to affinities between data and data representatives. The Kernel function encoded in SSVA is defined to rear to relational version and parameterized by supervisory information. Another point is the self-tuning of penalty coefficient and Kernel scale parameter to eliminate the need of searching parameter spaces. Experiments on real datasets demonstrate the performance and efficiency of SSVA. WB403-5 PID: 4968 Speeding up SVM in Test Phase: Application to Radar HRRP ATR Bo Chen, Xidian University, China Hongwei Liu, Xidian University, China Zheng Bao, Xidian University, China In this paper, a simple method is proposed to reduce the number of support vectors (SVs) in the decision function. Because in practice the embedded data just lie into a subspace of the kernel-induced space, F , we can search a set of basis vectors (BVs) to express all the SVs according to the geometrical structure, the number of which is less than that of SVs. The experimental results on real life and radar measured high-resolution range profile (HRRP) data sets show that our method can reduce the run-time complexity in SVM with the preservation of machine’s generalization, especially for the data of large correlation coefficients among input samples such as radar HRRP data.
WB404
71
WB404 Pattern Classification & Recognition II Section Chair: Thomas VILLMANN, University Leipzig WB404-1 PID: 5132 Prototype Based Classification Using Information Theoretic Learning Th. Villmann, University Leipzig, Germany B.Hammer, Clausthal University of Technology, Germany F.-M.Schleif, University Leipzig, Germany T. Geweniger, University Leipzig, Germany T. Fischer, University Leipzig, Germany M.Cottrell, University Paris I Sorbonne-Panth¨¦on, France In this article we extend the (recently published) unsupervised information theoretic vector quantization approach based on the Cauchy-Schwarz-divergence for matching data and prototype densities to supervised learning and classification. In particular, first we generalize the unsupervised method to more general metrics instead of the Euclidean, as it was used in the original algorithm. Thereafter, we extend the model to a supervised learning method resulting in a fuzzy classification algorithm. Thereby, we allow fuzzy labels for both, data and prototypes. Finally, we transfer the idea of relevance learning for metric adaptation known from learning vector quantization to the new approach. WB404-2 PID: 5406 A Modal Symbolic Classifier for Interval Data Fabio C.D. Silva, Federal University of Pernambuco, Brasil Francisco de A.T. de Carvalho, Federal University of Pernambuco, Brasil Renata M.C.R. de Souza, Federal University of Pernambuco, Brasil Joyce Q. Silva, Federal University of Pernambuco, Brasil A modal symbolic classifier for interval data is presented. The proposed method needs a previous pre-processing step to transform interval symbolic data into modal symbolic data. The presented classifier has then as input a set of vectors of weights. In the learning step, each group is also described by a vector of weight distributions obtained through a generalization tool. The allocation step uses the squared Euclidean distance to compare two modal descriptions. To show the usefulness of this method, examples with synthetic symbolic data sets are considered. WB404-3 PID: 4607 Hough Transform Neural Network for Seismic Pattern Detection Kou-Yuan Huang, National Chiao Tung University, Taiwan Jiun-De You, National Chiao Tung University, Taiwan Kai-Ju Chen, National Chiao Tung University, Taiwan Hung-Lin Lai, National Chiao Tung University, Taiwan An-Jin Don, National Chiao Tung University, Taiwan Hough transform neural network is adopted to detect line pattern of direct wave and hyperbola pattern of reflection wave in a seismogram. The distance calculation from
WB404
72
point to hyperbola is calculated from the time difference. This calculation makes the parameter learning feasible. The neural network can calculate the total error for distance from point to patterns. The parameter learning rule is derived by gradient descent method to minimize the total error. Experimental results show that line and hyperbola can be detected in both simulated and real seismic data. The network can get a fast convergence. The detection results can improve the seismic interpretation. WB404-4 PID: 4369 Efficient Domain Action Classification Using Neural Networks Hyunjung Lee, Sogang University, Korea; Harksoo Kim, Kangwon National University; Jungyun Seo, Sogang University, Korea Speaker’s intentions can be represented into domain actions (domain- independent speech acts and domain-dependent concept sequences). Therefore, domain action classification is very useful to a dialogue system that should catch user’s intention in order to generate correct reaction. In this paper, we propose a neural network model to determine speech acts and concept sequences at the same time. To avoid biased learning problems, the proposed model uses low-level linguistic features and filters out uninformative features using X^2 statistic. In the experiment, the proposed model showed better performances than the previous work in speech act classification. Moreover, the proposed model showed meaningful results when the size of training corpus was small. Based on the experimental results, we believe that the proposed model will be more helpful to dialogue systems because it manages speech act classification and concept sequence classification at the same time. We also believe that the proposed model can alleviate sparse data problems in speech act classification.
WB405
73
WB405 Biomedical Applications I Section Chair: Rinku PANCHAL, Central Queensland University WB405-1 PID: 5200 Characterisation of Breast Abnormality Patterns in Digital Mammograms using Auto-Associator Neural Network Rinku Panchal, Central Queensland University, Australia Brijesh Verma, Central Queensland University, Australia Presence of mass in breast tissues is highly indicative of breast cancer. The research work investigates the significance of neural-association of mass type of breast abnormality patterns for benign and malignant class characterization using auto-associator neural network and original features. The characterized patterns are finally classified into benign and malignant classes using a classifier neural network. Grey-level based statistical features, BI-RADS features, patient age feature and subtlety value feature have been used in proposed research work. The proposed research technique attained a 94% testing classification rate with a 100% training classification rate on digital mammograms taken from the DDSM benchmark database. WB405-2 PID: 4574 Ovarian Cancer Prognosis by Hemostasis and Complementary Learning T. Z. Tan, Nanyang Technological University, Singapore G. S. Ng, Nanyang Technological University, Singapore C. Quek, Nanyang Technological University, Singapore Stephen C. L. Koh, National University of Singapore, Singapore Ovarian cancer is a major cause of deaths worldwide. As a result, women are not diagnosed until the cancer has advanced to later stages. Accurate prognosis is required to determine the suitable therapeutic decision. Since abnormalities of hemostasis and increased risk of thrombosis are observed in cancer patient, assay involving hemostatic parameters can be potential prognosis tool. Thus a biological brain-inspired Complementary Learning Fuzzy Neural Network (CLFNN) is proposed, to complement the hemostasis in ovarian cancer prognosis. Experimental results that demonstrate the confluence of hemostasis and CLFNN offers a promising prognosis tool. Apart from superior performance, CLFNN provides interpretable rules to facilitate validation and justification of the system. Besides, CLFNN can be used as a concept validation tool for ovarian cancer prognosis. WB405-3 PID: 4835 Multi-Class Cancer Classification with OVR-Support Vector Machines Selected by Naive Bayes Classifier Jin-Hyuk Hong, Yonsei University, Korea Sung-Bae Cho, Yonsei University, Korea Support vector machines (SVMs), originally designed for binary classification, have been applied for multi-class classification, where an effective fusion scheme is required for
WB405
74
combining outputs from them and producing a final result. In this work, we propose a novel method in which the SVMs are generated with the one-vs-rest (OVR) scheme and dynamically organized by the naive Bayes classifiers (NBs). This method might break the ties that frequently occur when working with multi-class classification systems with OVR SVMs. More specifically, we use the Pearson correlation measure to select informative genes and reduce the dimensionality of gene expression profiles when constructing the NBs. The proposed method has been validated on GCM cancer dataset consisting of 14 types of tumors with 16,063 gene expression levels and produced higher accuracy than other methods. WB405-4 PID: 4558 Breast Cancer Diagnosis Using Neural-based Linear Fusion Strategies Yunfeng Wu, Beijing University of Posts and Telecommunications, China Cong Wang, Beijing University of Posts and Telecommunications, China S.C. Ng, The Open University of Hong Kong, Hong Kong Anant Madabhushi, Rutgers University, USA Yixin Zhong, Beijing University of Posts and Telecommunications, China Breast cancer is one of the leading causes of mortality among women, and the early diagnosis is of significant clinical importance. In this paper, we describe several linear fusion strategies, in particular the Majority Vote, Simple Average, Weighted Average, and Perceptron Average, which are used to combine a group of component multilayer perceptrons with optimal architecture for the classification of breast lesions. In our experiments, we utilize the criteria of mean squared error, absolute classification error, relative error ratio, and Receiver Operating Characteristic (ROC) curve to concretely evaluate and com-pare the performances of the four fusion strategies. The experimental results demonstrate that the Weighted Average and Perceptron Average strategies can achieve better diagnostic performance compared to the Majority Vote and Simple Average methods. WB405-5 PID: 5181 A Quantitative Diagnostic Method Based on Bayesian Networks in Traditional Chinese Medicine Huiyan Wang, Zhejiang Gongshang University, China Jie Wang, China Academy of Traditional Chinese Medicine, China Traditional Chinese Medicine (TCM) is one of the most important complementary and alternative medicines. Due to the subjectivity and fuzziness of diagnosis in TCM, quantitative model or methods are needed to facilitate the popularization of TCM. In this article, a novel quantitative method for syndrome differentiation based on BNs is proposed. First the symptoms are selected by a novel mutual information based symptom selection algorithm (MISS) and then the mapping relationships between the selected symptoms and key elements are constructed. Finally, the corresponding syndromes are output by combining the key elements. The results show that the diagnostic model obtains relative reliable predictions of syndrome, and its average predictive accuracy rate reach 91.68%, which testifies that the method we proposed is feasible and effective and can be expected to be useful in the modernization of TCM.
WB406
75
WB406 Forecasting & Prediction I Section Chair: Michael SMALL, Hong Kong Polytechnic University WB406-1 PID: 5433 Predicting Chaotic Time Series by Boosted Recurrent Neural Networks Mohammad Assaad, Universit¨¦ Fran?ois Rabelais de Tours, France Romuald Bon¨¦, Universit¨¦ Fran?ois Rabelais de Tours, France Hubert Cardot, Universit¨¦ Fran?ois Rabelais de Tours, France This paper discusses the use of a recent boosting algorithm for recurrent neural networks as a tool to model nonlinear dynamical systems. It combines a large number of RNNs, each of which is generated by training on a different set of examples. This algorithm is based on the boosting algorithm where difficult examples are concentrated on during the learning process. However, unlike the original algorithm, all examples available are taken into account. The ability of the method to internally encode useful information on the underlying process is illustrated by several experiments on well known chaotic processes. Our model is able to find an appropriate internal representation of the underlying process from the observation of a subset of the states variables. We obtain improved prediction performances. WB406-2 PID: 5124 Uncertainty in Mineral Prospectivity Prediction Pawalai Kraipeerapun, Murdoch University, Australia Chun Che Fung, Centre for Enterprise Collaboration in Innovative Systems, Australia Warick Brown, The University of Western Australia, Australia Kok Wai Wong, Murdoch University, Australia Tamas Gedeon, The Australian National University, Australia This paper presents an approach to the prediction of mineral prospectivity that provides an assessment of uncertainty. Two feed-forward backpropagation neural networks are used for the prediction. One network is used to predict degrees of favourability for deposit and another one is used to predict degrees of likelihood for barren, which is opposite to deposit. These two types of values are represented in the form of truth-membership and false-membership, respectively. Uncertainties of type error in the prediction of these two memberships are estimated using multidimensional interpolation. These two memberships and their uncertainties are combined to predict mineral deposit locations. The degree of uncertainty of type vagueness for each cell location is estimated and represented in the form of indeterminacy-membership value. The three memberships are then constituted into an interval neutrosophic set. Our approach improves classification performance compared to an existing technique applied only to the truth-membership value. WB406-3 PID: 5402 Thermal Deformation Prediction in Machine Tools By Using Neural Network Chuan-Wei Chang, Chung Yuan Christian University, Taiwan Yuan Kang, Chung Yuan Christian University, Taiwan
WB406
76
Yi-Wei Chen, Chung Yuan Christian University, Tung Nan Institute of Technology, Taiwan Ming-Hui Chu, Tung Nan Institute of Technology, Taiwan Yea-Ping Wang, Tung Nan Institute of Technology, Taiwan Thermal deformation is a nonlinear dynamic phenomenon and is one of the significant factors for the accuracy of machine tools. In this study, a dynamic feed-forward neural network model is built to predict the thermal deformation of machine tool. The temperatures and thermal deformations data at present and past sampling time interval are used train the proposed neural model. Thus, it can model dynamic and the nonlinear relationship between input and output data pairs. According to the comparison results, the proposed neural model can obtain better predictive accuracy than that of some other neural model. WB406-4 PID: 4888 Fuzzy Time Series Prediction Method Based on Fuzzy Recurrent Neural Network Rafik Aliev, Azerbaijan State Oil Academy, Azerbaijan Bijan Fazlollahi, Georgia State University, USA Rashad Aliev, Eastern Mediterranean University Babek Guirimov, Azerbaijan State Oil Academy One of the frequently used forecasting methods is the time series analysis. Time series analysis is based on the idea that past data can be used to predict the future data. Past data may contain imprecise and incomplete information coming from rapidly changing environment. Also the decisions made by the experts are subjective and rest on their individual competence. Therefore, it is more appropriate for the data to be presented by fuzzy numbers instead of crisp numbers. A weakness of traditional crisp time series forecasting methods is that they process only measurement based numerical information and cannot deal with the perception-based historical data represented by fuzzy numbers. Application of a fuzzy time series whose values are linguistic values, can overcome the mentioned weakness of traditional forecasting methods. In this paper we propose a fuzzy recurrent neural network (FRNN) based fuzzy time series forecasting method using genetic algorithm. The effectiveness of the proposed fuzzy time series forecasting method is tested on benchmark examples. WB406-5 PID: 5116 Research on Predictive Maintenance for Hydropower Plant Based on MAS and NN Weijin Jiang, China University of Geosciences, China Xiaohong Lin, Central South University, China As the development of the electrical power market, the maintenance automation has become an intrinsic need to increase the overall economic efficiency of hydropower plants. A Multi-Agent System (MAS) based model for the predictive maintenance system of hydropower plant within the framework of Intelligent Control-Maintenance-Management System (ICMMS) is proposed. All maintenance activities, form data collection through the recommendation of specific maintenance actions, are integrated into the system. In this model, the predictive maintenance system composed of four layers: Signal Collection, Data Processing, Diagnosis and Prognosis, and Maintenance
WB406
77
Decision-Making. Using this model a prototype of predictive maintenance for hydropower plant is established. Artificial Neural-Network (NN) is successfully applied to monitor, identify and diagnosis the dynamic performance of the prototype system online.
WB407
78
WB407 Fuzzy Systems I Section Chair: Frank LEUNG, Hong Kong Polytechnic University WB407-1 PID: 5532 A Look-ahead Fuzzy Back Propagation Network for Lot Output Time Series Prediction in a Wafer Fab Toly Chen, Feng Chia University, Taiwan Lot output time series is one of the most important time series data in a wafer fab (fabrication plant). Predicting the output time of every lot is therefore a critical task to the wafer fab. To further enhance the effectives and efficiency of wafer lot output time prediction, a look-ahead fuzzy back propagation network (FBPN) is constructed in this study with two advanced features: the future release plan of the fab is considered (look-ahead); expert opinions are incorporated. Production simulation is also applied in this study to generate test examples. According to experimental results, the prediction accuracy of the look-ahead FBPN was significantly better than those of four existing approaches: multiple-factor linear combination (MFLC), BPN, case-based reasoning (CBR), and FBPN without look-ahead, by achieving a 12%~37% (and an average of 19%) reduction in the root-mean-squared-error (RMSE) over the comparison basis ¨C MFLC. WB407-2 PID: 5511 Extraction of fuzzy features for detecting brain activation from functional MR time-series Juan Zhou, Nanyang Technological University, Singapore Jagath C. Rajapakse, Nanyang Technological University, Singapore We propose methods to extract fuzzy features from fMR time-series in order to detect brain activation. Five discriminating features are automatically extracted from fMRI using a sequence of temporal-sliding-windows. A fuzzy model based on these features is first developed by gradient method training on a set of initial training data and then incrementally updated. The resulting fuzzy activation maps are then combined to provide a measure of strength of activation for each voxel in human brain; a two-way thresholding scheme is introduced to determine actual activated voxels. The method is tested on both synthetic and real fMRI datasets for functional activation detection, illustrating that it is less vulnerable to correlated noise and is able to adapt to different hemodynamic response functions across subjects through incremental learning. WB407-3 PID: 5042 An Advanced Design Methodology of Fuzzy Set-based Polynomial Neural Networks with the aid of Symbolic gene type Genetic Algorithms and Information Granulation Seok-Beom Roh, Wonkwang University, Korea Hyung-Soo Hwang, Wonkwang University, Korea Tae-Chon Ahn, Wonkwang University, Korea In this paper, we propose a new design methodology that adopts Information Granulation to the structure of fuzzy-neural networks called Fuzzy Set-based Polynomial Neural Networks (FSPNN). We find the optimal structure of the proposed model with the aid of
WB407
79
symbolic genetic algorithms which has symbolic gene type chromosomes. We are able to find information related to real system with Information Granulation through numerical data. Information Granules obtained from Information Granulation help us understand real system without the field expert. In Information Granulation, we use conventional Hard C-Means Clustering algorithm and proposed procedure that handle the apex of clusters using ‘Union’ and ‘Intersection’ operation. We use genetic algorithm to find optimal structure of the proposed networks. The proposed networks are based on GMDH algorithm that makes whole networks dynamically. In other words, FSPNN is built dynamically with symbolic genetic algorithms. Symbolic gene type has better characteristic than binary coding GAs from the size of solution space’s point of view. Symbolic genetic algorithms are capable of reducing the solution space more than conventional genetic algorithms with binary genetype chromosomes. The performance of genetically optimized FSPNN (gFSPNN) with aid of symbolic genetic algorithms is quantified through experimentation where we use a number of modeling benchmarks data which are already experimented with in fuzzy or neurofuzzy modeling. WB407-4 PID: 4814 A Hybrid Self-Learning Approach for Generating Fuzzy Inference Systems Yi Zhou, Intelligent Systems Center 50 Nanyang Drive, Singapore Meng Joo Er, Intelligent Systems Center 50 Nanyang Drive, Singapore In this paper, a novel hybrid self-learning approach termed Enhanced Dynamic Self-Generated Fuzzy Q-Learning (EDSGFQL) for automatically generating a Fuzzy Inference System (FIS) is presented. In the EDSGFQL approach, the structure of an FIS is generated via Reinforcement Learning (RL) while the centers of Membership Functions (MFs) are updated by an extended Self Organizing Map (SOM) algorithm. The proposed EDSGFQL methodology can automatically create, delete and adjust fuzzy rules without any priori knowledge. In the EDS- GFQL approach, fuzzy rules of an FIS are regarded as agents of the entire system and all of the rules are recruited, adjusted and terminated according to their contributions and participation. At the mean time, the centers of MFs are adjusted to move to the real centers in the sense of feature representation by the extended SOM approach. Comparative studies on a wall-following task by a mobile robot have been done for the proposed EDSGFQL approach and other current methodologies and the demonstration results show that the proposed EDSGFQL approach is superior.
WC402
80
Wednesday Session C, 4 October 2006
WC402 Cognitive Processing / Learning II Section Chair: Vanessa YAREMCHUK, University of Alberta WC402-1 PID: 5287 Binding Mechanism of Frequency-specific ITD and IID Information in Sound Localization of Barn Owl Hidetaka Morisawa, Univ. of Electro-Communications, Japan Daisuke Hirayama, Univ. of Electro-Communications, Japan Kazuhisa Fujita, Univ. of Electro-Communications, Japan Yoshiki Kashimori, Univ. of Electro-Communications, Japan Takeshi Kambara, Univ. of Electro-Communications, Japan Barn owls perform sound localization based on analyses of interaural differences in arrival time and intensity of sound. Two kinds of neural signals representing the interaural time difference (ITD) and interaural intensity difference (IID) are processed in parallel in anatomically separate pathway. The values of ITD and IID are largely changed depending on frequency components of sound. The neural circuits involved in sound localization must bind the frequency-specific ITD and IID information to determine a spatial direction of sound source. However, little is known about the binding mechanism. We present here a neural network model of ICc ls in which the signals representing ITD and IID are first combined with each other. It is shown using our model how the neural maps can be generated in ICc ls by the excitatory inputs from ICc core representing ITD and the inhibitory inputs from VLVps representing IID. We show also that ICx neuron detects a spatial direction of sound source by making synaptic connections with ICc ls neurons encoding ITD and IID information of the sound source, based on Hebbian learning. WC402-2 PID: 5240 Effect of Feedback Signals on Tuning Shifts of Subcortical Neurons in Echolocation of Bat Seiichi Hirooka, Univ. of Electro-Communications, Japan Kazuhisa Fujita, Univ. of Electro-Communications, Japan Yoshiki Kashimori, Univ. of Electro-Communications, Japan Most species of bats making echolocation use Doppler-shifted frequency of ultrasonic echo pulse to measure the velocity of target. The neural circuits for detecting the target velocity are specialized for fine-frequency analysis of the second harmonic constant frequency (CF2) component of Doppler-shifted echoes. To perform the fine-frequency analysis, the feedback signals from cortex to subcortical and peripheral areas are needed. The feedback signals are known to modulate the tuning property of subcortical neurons. However, it is not yet clear about the neural mechanism for the modulation of the tuning property. We present here a neural model for detecting Doppler-shifted frequency of echo sound reflecting from a target. We show that the model reproduce qualitatively the experimental results on the modulation of tuning shifts of subcortical neurons. We also
WC402
81
clarify the neural mechanism by which the tuning property is changed depending on the feedback connections between cortical and subcortical neurons. WC402-3 PID: 4771 Musical Judgments in Majoy Key Contexts by Artificial Neural Networks Vanessa Yaremchuk, University of Alberta, Canada Michael R. W. Dawson, University of Alberta, Canada In this paper, we describe an artificial neural network (ANN) that is trained to rate the “fit” of a probe tone in the musical context of a major scale, in accordance with data published by Krumhansl [1] for a similar task performed by human subjects. The internal structure of the trained network is interpreted and formal western musical concepts are uncovered, involving the circle of fifths and circles of major thirds. WC402-4 PID: 5542 Semantic activation and cortical areas related to the lexical ambiguity and idiomatic ambiguity Gisoon Yu, Korea University, Korea Choong-Myung Kim, Korea University, Korea Dong Hwee, Kim, Korea University, Korea Kichun Nam, Korea University, Korea The main goal of the present study was to examine the cerebral regions associated with the lexical and idiomatic ambiguity resolution. In our experimentation by fMRI, we specially refer to two goals. The first goal was to show the difference between the activation areas related to several kinds of ambiguities. The second goal was to reveal the difference of the activation areas related to ambiguity resolution. In result, there are activations in not only the left temporal lobe but also right hemisphere, more particularly the right frontal lobe and the right temporal lobe.
WC403
82
WC403 Special Session K - New Trends in Self-Organizing Maps I Section Chair: Vanessa YAREMCHUK, University of Alberta WC403-1 PID: 5587 Generalization of the self-organizing map: From artificial neural networks to artificial cortexes Tetsuo Furukawa, Kyushu Institute of Technology, Japan Kazuhiro Tokunaga, Kyushu Institute of Technology, Japan This paper presents a generalized framework of the self-organizing map (SOM) applicable to more extended data classes rather than vector data. To realize such generalization, modular structure is adopted; thus it is called a modular network SOM (mnSOM), in which each reference vector unit of the conventional SOM is replaced by a functional module. Since users can choose the functional module from any trainable architectures such as neural networks, the mnSOM has a lot of flexibility as well as high ability of data processing. In this paper, the essential idea is introduced first, and then its theory is described. WC403-2 PID: 5586 SOM of SOMs: An Extension of SOM from ‘Map’ to ‘Homotopy’ Tetsuo Furukawa, Kyushu Institute of Technology, Japan This paper aims to propose an extension of SOMs called an “SOM of SOMs,” or SOM2, in which the mapped objects are self-organizing maps themselves. In SOM2, each nodal unit of the conventional SOM is replaced by a function module of SOM. Therefore, SOM2 can be regarded as a variation of a modular network SOM (mnSOM). Since each child SOM module in SOM2 is trained to represent an individual map, the parent map in SOM2 generates a self-organizing map representing the continuous change of the child maps. Thus SOM2 is an extension of SOM that generates a ‘self-organizing homotopy’ rather than a map. This extension of SOM is easily generalized in the case of SOMn, such that “SOM3 as SOM of SOM2s”, corresponding to the n-th order of homotopy. In this paper, the algorithm of SOM2 is introduced, and some simulation results are reported. WC403-3 PID: 5588 Modular network SOM: Theory, algorithm and applications Kazuhiro Tokunaga, Kyushu Institute of Technology, Japan Tetsuo Furukawa, Kyushu Institute of Technology, Japan The modular network SOM (mnSOM) proposed by authors is an extension and generalization of a conventional SOM in which each nodal unit is replaced by a module such as a neural network. It is expected that the mnSOM will extend the area of applications beyond that of a conventional SOM. We set out to establish the theory and algorithm of a mnSOM, and to apply it to several research topics, to create a fundamental technology that is generally usable only in expensive studies. In this paper, the theory and the algorithm of the mnSOM are reported; moreover, the results of applications of the mnSOM are presented.
WC403
83
WC403-4 PID: 5582 Task Segmentation in a Mobile Robot by mnSOM: A New Approach to Training Expert Modules Muhammad Aziz Muslim, Kyushu Institute of Technology, Japan Masumi Ishikawa, Kyushu Institute of Technology, Japan Tetsuo Furukawa, Kyushu Institute of Technology, Japan Proposed is a new approach to task segmentation in a mobile robot by a modular network SOM (mnSOM). The standard mnSOM supposes that data are divided into classes with labels a priori. In a mobile robot, however, only a sequence of data without segmentation is available. Hence, we propose to decompose it into many subsequences, supposing that a class label does not change within a subsequence. Accordingly, training of mnSOM is done for each subsequence in contrast to that for each class in the standard mnSOM. The resulting mnSOM demonstrates good segmentation performance of 94.05% for a novel dataset. WC403-5 PID: 4457 Improving the Generalization of Fisherface by Training Class Selection Using SOM2 Jiayan Jiang, Fudan University, China Liming Zhang, Fudan University, China Tetsuo Furukawa, 2Kyushu Institute of Technology, Japan Fisherface is a popular subspace algorithm used in face-recognition, and is commonly believed superior to another technique, Eigenface, due to its attempt to maximize the separability of training classes. However, the obtained discriminating subspace of the training set may not easily extend to unseen classes, as in the case of enrollment of new subjects. In this paper, we select some representative classes for Fisherface training using a recently proposed neural network architecture SOM2. The experiments on ORL face database show that the proposed method can effectively reduce the performance variance and improve the generalization of Fisherface.
WC404
84
WC404 Image Processing I Section Chair: Benedicte BASCLE, France Telecom WC404-1 PID: 5423 A statistical approach for learning invariants: application to image color correction and learning invariants to illumination B. Bascle, Orange / France Telecom R & D, France O. Bernier, Orange / France Telecom R & D, France V. Lemaire, Orange / France Telecom R & D, France This paper presents a new approach for automatic image color correction, based on statistical learning. The method both parameterizes color independently of illumination and corrects color for changes of illumination. The motivation for using a learning approach is to deal with changes of lighting typical of indoor environments such as home and office. The method is based on learning color invariants using a modified multi-layer perceptron (MLP). The MLP is odd-layered. The middle layer includes two neurons which estimate two color invariants and one input neuron which takes in the luminance desired in output of the MLP. The advantage of the modified MLP over a classical MLP is better performance and the estimation of invariants to illumination. The trained modified MLP can be applied using look-up tables (LUTs), yielding very fast processing. Results illustrate the approach. WC404-2 PID: 5093 Limited Recurrent Neural Network for Superresolution Image Reconstruction Yan Zhang, Zhengzhou Institute of Surveying and Mapping, China Qing Xu, Zhengzhou Institute of Surveying and Mapping, China; Tao Wang, Zhengzhou Institute of Surveying and Mapping, China Lei Sun, Zhengzhou Institute of Surveying and Mapping, China The paper proposes a new method for image resolution enhancement from multiple images using the limited recurrent neural network (LRNN) approach, which is a set of collectively operating feed-forward neural networks. In the limited recurrent networks, information about past outputs is fed back through recurrent connections of output units and mixed with the input nodes flowing into the network input as external input nodes. Thus, experience about past search is utilized, which enables LRNN to be capable of both learning and searching the optimal solution for optimization problems in the solution space. Estimates computed from a low-resolution (LR) simulation image sequence and an actual video film sequence show dramatic visual and quantitative improvements over bilinear interpolation, and equivalent performance to that of the frequency domain approach. WC404-3 PID: 4897 Active contour with neural networks-based information fusion kernel Xiongcai Cai, The University of New South Wales, Australia, National ICT Australia, Australia
WC404
85
Arcot Sowmya, The University of New South Wales, Australia, UNSW Asia, Singapore, National ICT Australia, Australia This paper proposes a novel active contour model for image object recognition using neural networks as a dynamic information fusion kernel. It first learns feature fusion strategies from training data by searching for an optimal fusion model at each marching step of the active contour model. A recurrent neural network is then employed to learn the fusion strategy knowledge. The learned knowledge is then applied to guide another linear neural network to fuse the features, which determine the marching procedures of an active contour model for object recognition. We test our model on both artificial and real image data sets and compare the results to those of a standard active model, with promising outcomes. WC404-4 PID: 5398 A split-and-merge method with ranking and selection for polygonal approximation of digital curve Bin Wang, Fudan University, China Chaojian Shi, Fudan University, Shanghai Maritime University, China How to use a polygon with the fewest possible sides to approximate a shape boundary is an important issue in pattern recognition and image processing. A novel split-and-merge technique (SMT) is proposed. SMT starts with an initial shape boundary segmentation, split and merge are then alternately done against the shape boundary. The procedure is halted when the pre-specified iteration number is achieved. For increasing stability of SMT and improving its robustness to the initial segmentation, a ranking-selection scheme is utilized to choose the splitting and merging points. The experimental results show its superiority. WC404-5 PID: 5050 MPEG Video Traffic Modeling and Classification using Fuzzy C-Means Algorithm with Divergence-based Kernel Dong-Chul Park, Myong Ji University, Korea Chung Nguyen Tran, Myong Ji University, Korea Yunsik Lee, , Korea Electronics Tech. Inst., Korea A modeling and classification model for MPEG video traffic data using a Fuzzy C-Means algorithm with a Divergence-based Kernel (FCMDK) for clustering GPDF data is proposed in this paper. The FCMDK is based on the Fuzzy C-Means clustering algorithm and thus exploits advantageous features of fuzzy clustering techniques. To further improve classification accuracies and deal with nonlinear data, the input data is projected into a feature space of a higher dimensionality. Consequently, nonlinear problems existing in the input space can be solved linearly in the feature space. The divergence-based kernel method adopted in the FCMDK employs a divergence measure between two probability distributions for its similarity measure. By adopting the divergence-based kernel method for probability data, the FCMDK can not only utilize advantageous features of the kernel method but also exploit the statistical nature of the input data. Experiments and results on several MPEG video traffic data sets demonstrate that the classification model employing the FCMDK for clustering GPDF data can archive
WC404
86
improvements of 28.19% and 34.60% in terms of False Alarm Rate (FAR) over the models using the conventional k-means and SOM algorithms, respectively.
WC405
87
WC405 Special Session A-Computational Intelligence in Economics and Finance Section Chair: Shu-Heng CHEN, National Chengchi University; Nicolas NAVET, INRIA WC405-1 PID: 5005 Combining Time-Scale Feature Extractions with SVMs for Stock Index Forecasting Shian-Chang Huang, National Changhua University of Education, China Hsing-Wen Wang, National Changhua University of Education, China Support vector machine (SVM) has appeared as a powerful tool for time series forecasting and demonstrated better performance over other methods, such as neural networks or ARIMA-based models. This paper proposes a hybrid model to combine time-scale feature extractions with SVM models for stock index forecasting. The time series of explanatory variables are decomposed by the wavelet basis, and the extracted time-scale features serve as inputs of a SVM model to perform the nonparametric forecasting. Compared with pure SVM models or traditional GARCH models, the performance of the new method is the best. The results of this study can help investors for controlling and reducing their risks in international investments. WC405-2 PID: 5520 A Double-Phase Genetic Optimization Algorithm for Portfolio Selection Kin Keung Lai, Hunan University, China Lean Yu, City University of Hong Kong, Hong Kong Shouyang Wang, Hunan University, China Chengxiong Zhou, Chinese Academy of Sciences, China In this study, a double-stage genetic optimization algorithm is proposed for portfolio selection. In the first stage, a genetic algorithm is used to identify good quality assets in terms of asset ranking. In the second stage, investment allocation in the selected good quality assets is optimized using a genetic algorithm based on Markowitz’s theory. Through the two-stage genetic optimization process, an optimal portfolio can be determined. Experimental results reveal that the proposed double-stage genetic optimization algorithm for portfolio selection provides a very feasible and useful tool to assist the investors in planning their investment strategy and constructing their portfolio. WC405-3 PID: 5559 Pretests for genetic-programming evolved trading programs: “zero-intelligence” strategies and lottery trading Shu-Heng Chen, National Chengchi University, Taiwan Nicolas Navet, National Chengchi University, Campus-Scientifique, Taiwan Over the last decade, numerous papers have investigated the use of GP for creating financial trading strategies. Typically in the literature results are inconclusive but the investigators always suggest the possibility of further improvements, leaving the conclusion regarding the effectiveness of GP undecided. In this paper, we discuss a series of pretests, based on several variants of random search, aiming at giving more clear-cut answers on whether a GP scheme, or any other machine-learning technique, can be
WC405
88
effective with the training data at hand. The analysis is illustrated with GP-evolved strategies for three stock exchanges exhibiting different trends. WC405-4 PID: 5570 Intelligence-Based Model to Timing Problem of Resources Exploration in the Behavior of Firm Hsiu Fen Tsai, Shu-Te University, Taiwan Bao Rong Chang, National Taitung University, Taiwan We have insight into the importance of resource exploration derived from the quest for sustaining competitive advantage as well as the growth of the firm, which are well-explicated in the resources-based view. However, we really do not know when the firm will seriously commit to this kind of activities. Therefore, this study proposes an intelligence-based model using quantum minimization (QM) to tune a composite model of adaptive neuron-fuzzy inference system (ANFIS) and nonlinear generalized autoregressive conditional heteroscedasticity (NGARCH) such that it constitutes the relationship among five indicators, the growth rate of long-term investment, the firm size, the return on total asset, the return on common equity, and the return on sales. In particularly, this proposed approach outperforms several typical methods such as autoregressive moving-average regression (ARMAX), back-propagation neural network (BPNN), or adaptive support vector regression (ASVR) for this timing problem in term of comparing their achievement and the goodness of fit. Consequently, the preceding methods involved in this problem truly explain the timing of resources exploration in the behavior of firm. Meanwhile, the performance summary among methods is compared quantitatively. WC405-5 PID: 5579 Exchange Options Pricing with Evolutionary Neural-based Fuzzy Inference Systems Hsing-Wen Wang, National Changhua University of Education, Taiwan Since 1973, Fisher Black and Myron Scholes derived the BSM that provides a short-cut pricing method for options. And then, the option markets and earlier studies take the BSM as the practical model and develop more and more prospering. However, BSM based on many assumptions and constrains such that the price derived with this model was incorrect while compared with the practical prices of the market. Over the past years, many researches in computational intelligence areas reveal that the artificial neural networks (ANNs) pertain excellent learning, high speed computing capabilities, fault-tolerance abilities and the capability of processing non-linear problems to overcome the drawbacks derived from BSM. On the other hand, the adaptive structure neural networks integrated with fuzzy inference systems optimized using Extended-Kalman predictor and modified backpropagation algorithms while considering the fuzzy environments are proposed by Jang (1993). We employ the proposed options pricing model through enhanced neural-fuzzy-based inference systems (ENFIS), whose initial parameters of premise universe can be adjusted systematically by enhanced fuzzy c-means clustering method (EFCM) and programming initially consequence universe with genetic algorithms (GAs) in options pricing and then compared with the BSM. The evidence from empirical studies is using the Deutsche Mark foreign exchange of the Chicago Mercantile Exchange (CME). Results generated by the ENFIS pricing model would be compared with the
WC405
89
original BSM using various volatility strategies through pricing error measurement and interpret capability in the research period from 1990 to 1992. The results show that the ENFIS pricing model is superior to the BSM no matter in error degree, variation degree or in interpretation capability.
WC406
90
WC406 Robotics and Control I Section Chair: Wai-keung FUNG, University of Manitoba WC406-1 PID: 4823 Constrained Multi-variable Generalized Predictive Control Using a Dual Neural Network Long Cheng, The Chinese Academy of Sciences, China Zeng-Guang Hou, The Chinese Academy of Sciences, China Min Tan, The Chinese Academy of Sciences, China A dual neural network is applied to incorporate with the multi-variable generalized predictive control (MGPC) algorithm. Constraints of the process system can be taken into account and optimum values of future control signals can be obtained exactly. Compared to other control schemes, this approach is simpler, faster, and easier to be implemented. Hence it is suitable for industrial applications with the real-time computing requirement. A simulation example is presented to demonstrate its efficiency. WC406-2 PID: 4896 Turbulence Encountered Landing Control Using Hybrid Intelligent System Jih-Gau Juang, National Taiwan Ocean University, Taiwan Hou-Kai Chiou, National Taiwan Ocean University, Taiwan During a flight, take-off and landing are the most difficult operations in regard to safety issues. Aircraft pilots must not only be acquainted with the operation of instrument boards but also need flight sensitivity to the ever-changing environment, especially in the landing phase when turbulence is encountered. If the flight conditions are beyond the preset envelope, the automatic landing system (ALS) is disabled and the pilot takes over. An inexperienced pilot may not be able to guide the aircraft to a safe landing at the airport. This paper proposes an intelligent aircraft automatic landing controller that uses recurrent neural network (RNN) controller with genetic algorithm (GA) to improve the performance of conventional ALS and guide the aircraft to a safe landing. WC406-3 PID: 5296 An AND-OR Fuzzy Neural Network Ship Controller Design Jianghua Sui, Dalian Maritime University, China Guang Ren, Dalian Maritime University, China In this paper, an AND-OR fuzzy neural network (AND-OR FNN) and a piecewise optimization approach are proposed. The in-degree of neuron and the connectivity of layer are firstly defined and Zadeh’s operators are employed in order to infer the symbolic expression of every layer, the equivalent is proved between the architecture of AND-OR FNN and fuzzy weighted Mamdani inference. The main superiority is shown not only in reducing the input space, but also auto-extracting the rule base. The optimization procedure consists of GA (Genetic Algorithm) and PA (Pruning Algorithm); the AND-OR FNN ship controller system is designed based on input-output data to validate this method. Simulating results demonstrate that the number of rule base is decreased
WC406
91
remarkably and the performance is good, illustrate the approach is practicable, simple and effective. WC406-4 PID: 4266 Hierarchical Multiple Models Neural Network Decoupling Controller for a Nonlinear System Xin Wang, Shanghai Jiao Tong University, China Hui Yang, Shanghai Jiao Tong University, China For a nonlinear discrete-time Multi-Input Multi-Output (MIMO) system, a Hierarchical Multiple Models Neural Network Decoupling Controller (HMMNNDC) is designed in this paper. Firstly, the nonlinear system’s working area is partitioned into several sub-regions by use of a Self-Organizing Map (SOM) Neural Network (NN). In each sub-region, around every equilibrium point, the nonlinear system can be expanded into a linear term and a nonlinear term. Therefore the linear term is identified by a BP NN trained offline while the nonlinear term by a BP NN trained online. So these two BP NNs compose one system model. At each instant, the best sub-region is selected out by the use of the SOM NN and the corresponding multiple models set is derived. According to the switching index, the best model in the above model set is chosen as the system model. To realize decoupling control, the nonlinear term and the interaction of the system are viewed as measurable disturbance and eliminated using feedforward strategy. The simulation example shows that the better system response can be got comparing with the conventional NN decoupling control method.
WC407
92
WC407 Data Preprocessing/Feature Selection I Section Chair: Takashi MINOAHRA, Takushoku University WC407-1 PID: 5270 Fault Tolerant Training of Neural Networks for Learning Vector Quantization Takashi MINOHARA, Takushoku University, Japan The learning vector quantization (LVQ) is a model of neural networks, and it is used for complex pattern classifications in which typical feedforward networks don’t give a good performance. Fault tolerance is an important feature in the neural networks, when they are used for critical application. Many methods for enhancing the fault tolerance of neural networks have been proposed, but most of them are for feedforward networks. There are scarcely any methods for fault tolerance of LVQ neural networks. In this paper, we proposed a dependability measure for the LVQ neural networks, and then we presented two idea, the border emphasis and the encouragement of coupling, to improve the learning algorithm for increasing dependability. The experiment result shows that the proposed algorithm trains networks so that they can achieve high dependability. WC407-2 PID: 5415 Clustering with a Semantic Criterion Based on Dimensionality Analysis Wenye Li, The Chinese University of Hong Kong, Hong Kong Kin-Hong Lee, The Chinese University of Hong Kong, Hong Kong Kwong-Sak Leung, The Chinese University of Hong Kong, Hong Kong Considering data processing problems from a geometric point of view, previous work has shown that the intrinsic dimension of the data could have some semantics. In this paper, we start from the consideration of this inherent topology property and propose the usage of such a semantic criterion for clustering. The corresponding learning algorithms are provided. Theoretical justification and analysis of the algorithms are shown. Promising results are reported by the experiments that generally fail with conventional clustering algorithms. WC407-3 PID: 4931 Message-passing for inference and optimization of real variables on sparse graphs K. Y. Michael Wong, Hong Kong University of Science and Technology, Hong Kong C. H. Yeung, Hong Kong University of Science and Technology, Hong Kong David Saad, Aston University, UK The inference and optimization in sparse graphs with real variables is studied using methods of statistical mechanics. Efficient distributed algorithms for the resource allocation problem are devised. Numerical simulations show excellent performance and full agreement with the theoretical results. WC407-4 PID: 5063 Analysis and Insights into the Variable Selection Problem
WC407
93
Amir F. Atiya, Cairo University, Egypt In many large applications a large number of input variables are initially available, and a subset selection step is needed to select the best few to be used in the subsequent classification or regression step. The designer initially screens the inputs for the ones that have good predictive ability and that are not too much correlated with the other selected inputs. In this paper, we study how the predictive ability of the inputs, viewed individually, reflects on the performance of the group (i.e. what are the chances that as a group they perform well). We also study the effect of “irrelevant” inputs. We develop a formula for the distribution of the change in error due to adding an irrelevant input. This can be a useful reference. We also study the role of correlations and their effect on group performance. To study these issues, we first perform a theoretical analysis for the case of linear regression problems. We then follow with an empirical study for nonlinear regression models such as neural networks. WC407-5 PID: 4620 Dimensionality Reduction of Protein Mass Spectrometry Data using Random Projection Chen Change Loy, Grid Computing and Bioinformatics Lab, Malaysia Weng Kin Lai, Grid Computing and Bioinformatics Lab, Malaysia Chee Peng Lim, University of Science Malaysia, Malaysia Protein mass spectrometry (MS) pattern recognition has recently emerged as a new method for cancer diagnosis. Unfortunately, classification performance may degrade owing to the enormously high dimensionality of the data. This paper investigates the use of Random Projection in protein MS data dimensionality reduction. The effectiveness of Random Projection (RP) is analyzed and compared against Principal Component Analysis (PCA) by using three classification algorithms, namely Support Vector Machine, Feed-forward Neural Networks and K-Nearest Neighbour. Three real-world cancer data sets are employed to evaluate the performances of RP and PCA. Through the investigations, RP method demonstrated better or at least comparable classification performance as PCA if the dimensionality of the projection matrix is sufficiently large. This paper also explores the use of RP as a pre-processing step prior to PCA. The results show that without sacrificing classification accuracy, performing RP prior to PCA significantly improves the computational time.
WP
94
Wednesday Poster Session I, 4 October 2006
WP-1 PID: 4249 A divide-and-conquer approach to the Pairwise Opposite Class-Nearest Neighbor (POC-NN) algorithm for regression problem\ Thanapant Raicharoen, Chulalongkorn University, Thailand Chidchanok Lursinsap, Chulalongkorn University, Thailand Frank Lin, University of Maryland Eastern Shore, Thailand This paper presents a method for regression problem based on divide-and-conquer approach to the selection of a set of prototypes from the training set for the nearest neighbor rule. This method aims at detecting and eliminating redundancies in a given data set while preserving the significant data. A reduced prototype set contains Pairwise Opposite Class-Nearest Neighbor (POC-NN) prototypes which are used instead of the whole given data. Before finding POC-NN prototypes, all sampling data have to be separated into two classes by using the criteria through odd and even sampling number of data, then POC-NN prototypes are obtained by iterative separation and analysis of the training data into two regions until each region is correctly grouped and classified. The separability is determined by the POC-NN prototypes essential to define the function approximator for local sampling data locating near these POC-NN prototypes. Experiments and results reported showed the effectiveness of this technique and its performance in both accuracy and prototype rate to those obtained by classical nearest neighbor techniques. WP-2 PID: 4305 Delay-Dependent and Delay-Independent Stability Conditions of Delayed Cellular Neural Networks Wudai Liao, Zhongyuan University of Technology, China Dongyun Wang, Zhongyuan University of Technology, China Yulin Xu, Zhongyuan University of Technology, China Xiaoxin Liao, Huazhong University of Science and Technology, China By using the saturation linearity of the output functions of neurons in cellular neural networks, and by adopting the method of decomposing the state space to sub-regions, the mathematical equations of delayed cellular neural networks are rewritten to be the form of linear differential difference equations in the neighbourhood of each equilibrium, which is an interior point of some sub-region. Based on this linear form and by using the stability theory of linear differential difference equations and the tool of M-matrix, delay-dependent and delay-independent stability algebraic criteria are obtained. All results obtained in this paper need only to compute the eigenvalues of some matrices or to examine the matrices to be M-matrix or to verify some inequalities to be held. WP-3 PID: 4417 A RBFN Based Cogging Force Estimator of PMLSM with Q<1 Structure Bo Shao, Zhejiang University, China Zhitong Cao, Zhejiang University, China Yuetong Xu, Zhejiang University, China
WP
95
The cogging force has great impact to the efficiency of permanent magnetic liner synchronous motor (PMLSM) especially at high precision and low speed. This paper presents a cogging force estimator based on radical basis functional network (RBFN) by accelerated fuzzy c-means algorithm. Comparing to the estimator based on back propagation neural network (BPNN) with momentum method, the novel estimator increases the clustering of NN by boosting learning rates. Simulation results show the fractional slot with q<1 structure effectively depresses cogging force in PMLSM. Experiments prove that the estimator has high accuracy and efficiency. The novel estimator achieves demand of agility design and gives reference for structural parameters selection in PMLSM. WP-4 PID: 4559 An Extended Model on Self-Organizing Map Shuzhong Yang, Beijing Jiaotong University, China Siwei Luo, Beijing Jiaotong University, China Jianyu Li, Communication University of China, China In this paper, we present an extended self-organizing map. It keeps full connectivity between adjacent layers but adds new virtual connections between neurons of competitive layer so that the structure of competitive layer can be regarded as a graph and can be expressed by an adjacent matrix. Thus, the conventional SOMs can be regarded as special cases of the extended model. Then we can evolve the graph into arbitrary topology such as small world graph and random graph. After evolution we can obtain arbitrary nonlinear neighborhood kernel of neurons and the obtained topology of competitive layer is expected to simulate the distribution of input samples. The experimental results show that the new extended model has better performance in speed and self-organization than conventional ones. WP-5 PID: 4596 Analysis of Dynamics of Cultured Neuronal Networks Using I & F Model Akio KAWANA, Takushoku University, Japan Shinya TSUZUKI, Takushoku University, Japan Osamu WATANABE, Takushoku University, Japan Experimentally, spontaneous activity of the developing cultured neuronal networks of cortical neurons is known to show three different types of dynamics according to their developmental stages. As these changes of dynamics play an important role in development of the networks, it is necessary to explore the mechanism of these changes of dynamics for understanding how the networks develop. However, little is known concerning the mechanism from the experiments because of difficulties in controlling experimental conditions. The simulation of electrical activity of neuronal networks formed by simple integrated and re neurons provides a useful method for exploring such changes of dynamics. The simulation suggests that these changes of dynamics were mainly due to development of connectivity of synapses, changes of relative refractory period and development of inhibitory synapses. WP-6
WP
96
PID: 4608 A Novel Speech Processing Algorithm for Cochlear Implant Based on Selective Fundamental Frequency Control Tian Guan, Tsinghua University, China Qin Gong, Tsinghua University, China Datian Ye, Tsinghua University, China The speech recognition ability of most CI users in noisy environment remains quite poor, especially for those who speak tonal language, such as Mandarin. Based on the results of the acoustic research on Mandarin, a novel algorithm using fundamental frequency was proposed, adding the principle of frequency bands selection. Using this principle F0 was encoded only in the selected frequency bands, which was confirmed effectively in acoustic simulation experiments with white noise or mixed speech environment. Compared with the traditional algorithm which only transmitted envelope cues, this novel strategy achieved remarkable improvement no matter adopting Mandarin vowels, words, sentences or mixed sentences. What’s more, the amount of transmission in this algorithm decreased to 62.5% compared with similar algorithms which transmit fundamental frequency in full channels. WP-7 PID: 4652 A New Approach for the Short-Term Load Forecasting with Autoregressive and Artificial Neural Network Models U. Basaran Filik, Anadolu University, Turkey M. Kurban, Anadolu University, Turkey In this paper, a new approach for the short-term load forecasting with autoregressive (AR) and artificial neural network (ANN) models is introduced and also applied to the power system of Turkey by using the consumption values of electrical energy for three months, January, February, and March in 2002. Therefore, the load forecasting for the next day using AR and ANN models is performed separately, then results of the AR analysis is used for the input of different ANN models which are Feed Forward Back Propagation and Cascade Forward Back Propagation Models as a new approach. All of them are compared with each other. When the whole weeks are examined in point of the energy consumption, it is seen that Sundays are different from other six days in the weeks. Because of this, the values for the past six days except Sunday are used for the load forecasting of the next day. Although the network is constructed with 6-neuron input layer and 1-neuron output layer in the application used only ANN model, there are 7-neuron input layer and 1-neuron output layer in the structure of the new ANN model which forecasts the better using the results of AR model as the plus input values. WP-8 PID: 4663 Collective information-theoretic competitive learning Ryotaro Kamimura, Information Science Laboratory, Japan Fumihiko Yoshida, Tokai University, Japan In this paper, we try to show that the simple collection of competitive units can show some emergent property such as improved generalization performance. We have so far defined information-theoretic competitive learning with respect to individual competitive
WP
97
units. As information is increased, one competitive unit tends to win the competition. This means that competitive learning can be described as a process of information maximization. However, in living systems, a large number of neurons behave collectively. Thus, it is urgently needed to introduce collective property in information-theoretic competitive learning. In this context, we try to treat several competitive units as one unit, that is, one collective unit. Then, we try to maximize information content not in individual competitive units but in collective competitive units. We applied the method to an artificial data, the well-known voting attitude problem and cabinet approval rating estimation. In all cases, we successfully demonstrated that improved generalization could be obtained. WP-9 PID: 4712 Neurocomputing for Minimizing Energy Consumption of Real-Time Operating System in the System-on-a-chip Bing Guo, SiChuan University, China Dianhui Wang, Yan Shen, La Trobe University, Australia Zhishu Li, University of Electronic Science & Technology of China, China The RTOS (Real-Time Operating System) is a critical component in the SoC (System-on-a-Chip), which consumes the dominant part of total system energy. A RTOS system-level power optimization approach based on hardware- software partitioning (RTOS-Power partitioning) can significantly minimize the energy consumption of a SoC. This paper presents a new model for RTOS-Power partitioning, which helps in understanding the essence of the RTOS-Power partitioning techniques. A discrete Hopfield neural network approach for implementing the RTOS-Power partitioning is proposed, where a novel energy function, operating equation and coefficients of the neural network are redefined. Simulations are carried out with comparison to other optimization techniques. Experimental results demonstrate that the proposed method can achieve higher energy savings up to 60% at relatively low costs. WP-10 PID: 4799 A New Mechanism on Brain Information Processing—Energy Coding Rubin Wang, East China University of Science and Technology, China Zhikang Zhang, East China University of Science and Technology, China According to the experimental result of signal transmission with energetic demand tightly coupled to the information coding in cerebral cortex and electric structural property in neuronal activities, we present a brand-new scientific theory with mechanism of brain information processing. According to the new theory, we discover that neural coding under action of stimulation in brain is complete with way of energy coding. Due to energy coding to be able to reveal mechanism of brain information processing in physical essence, we can not only finely reproduce various experimental results of neuro-electrophysiology, but also quantitatively explain the experimental results from neuroscientists at Yale University in recently. Due to theory of energy coding to bridge the gap between functioning connection of biological neural network and energetic consumption, we can estimate that the new theory possesses very important sense for quantitative research of cognitive functioning.
WP
98
WP-11 PID: 4974 Commitment and Typicality Measurements for the Self-Organizing Map Zhe Li, Clark University, United States of America J. Ronald Eastman, Clark University, United States of America In this paper, we propose two non-parametric algorithms for the SOM to provide soft classification outputs. These algorithms, which are labeling-frequency-based and are called SOM Commitment (SOM-C) and SOM Typicality (SOM-T), expressing in the first case the degree of commitment the classifier has for each class for a specific pixel and in the second case, how typical that pixel’s reflectances are of the ones upon which the classifier was trained for each class. To evaluate the two proposed algorithms, soft classifications of a SPOT HRV image were undertaken. A Bayesian posterior probability soft classifier and a Mahalanobis typicality soft classifier were also used as a comparison. Principal Components Analysis (PCA) was used to explore the relationship between these measures. Results indicate that great similarities exist between the SOM-C and a parametric Bayesian posterior probability classifier, and between the SOM-T and a Mahalanobis typicality classifier. WP-12 PID: 5006 Adaptively incremental Self-organizing Isometric* Embedding Hou Yuexian, Tianjin University, China Gong Kefei, Tianjin University, China He Pilian, Tianjin University, China In this paper, we propose an adaptive incremental nonlinear dimensionality reduction algorithm for data stream in adaptive Self-organizing Isometric Embedding [1][3] framework. Assuming that each sampling point of underlying manifold and its adaptive neighbors [3] can preserve the principal directions of the regions that they reside on, our algorithm need only update the geodesic distances between anchors and all the other points, as well as distances between neighbors of incremental points and all the other points when a new point arrives. Under the above assumption, our algorithms can realize an approximate linear time complexity embedding of incremental points and effectively tradeoff embedding precision and time cost. WP-13 PID: 5027 Hierarchical and Interpretable Connectionist Structure Generation from Data Waratt Rattasiri, The University of Melbourne, Australia Saman K. Halgamuge, The University of Melbourne, Australia Nalin Wickramarachchi, The University of Melbourne, Australia In this paper, a self-generating hierarchical fuzzy system, namely Hierarchical Neuro-Fuzzy System (HiNeFS), is presented. While Hierarchical Neuro-Fuzzy System relies on the structure of the initial training network, it utilizes the newly proposed analytical criterions to identify the qualified rule relevant nodes based on their history of activities and behaviors which reflect their levels of involvement and contribution during the learning process. The proposed criterions have proved to be able to effectively reduce the size of the network, hence the computational complexity, and the resulting hierarchical
WP
99
network has demonstrated its capability in efficiently performing its task. To verify the proposed system performance, it is tested against two well-known benchmark datasets whose results are provided and discussed. WP-14 PID: 5028 Design and Implementation of Human Location and Motion Recognition System based on LSVM Juyeon Lee, Sejong University, Korea Jonghwa Choi, Sejong University, Korea Dongkyoo Shin, Sejong University, Korea Dongil Shin, Sejong University, Korea A smart home is an invisible home agent that collects home contexts that happen in a home, and it provides an automatic home service for a human. A smart environment must be able to offer a spontaneous home service that selects the most suitable of the human’s information. A home agent collects various home contexts, and predicts a home service for the human through recognition of a context pattern. The most important context is human context (location and motion) in all contexts. This paper presents the human location and motion recognition system that is based human’s absolute coordinates and relative co-ordinates in a home. We present the architecture of the human location and motion recognition system, and explain algorithms of all components. We present a processing algorithm that is applied to three kinds of image for human’s location recognition, and apply a LSVM (linear support vector machine) to predict human motion. In our experiment, the prediction of the human location had an error range of 0.024 m, and the prediction of the human motion had a high ac-curacy of 88.5%. Since we tested with only one human (user), we are currently studying location and motion recognition for multi-users. WP-15 PID: 5097 Reinforcement Learning Approaches for Constrained MDPs Peter Geibel, University of Osnabr¨uck, Germany Most reinforcement learning approaches consider Markov decision processes (MDPs) with a single criterion to be maximized. In practical applications, however, we are often confronted with additional criteria, e.g. constraints on the energy or time consumed during solving the main task, which is expressed by the first criterion. In this article, we therefore consider Markov Decision Processes with two criteria, each defined as the expected value of an infinite horizon cumulative return. The second criterion is subject to an inequality constraint. We describe several dynamic programming and reinforcement learning approaches for solving such control problems, discuss their advantages and shortcomings, and present experimental results based on randomly generated MDPs. WP-16 PID: 5149 Alpha-Beta Bidirectional Associative Memories María Elena Acevedo-Mosqueda, Laboratorio de Inteligencia Artificial, Mexico Cornelio Yáñez-Márquez, Laboratorio de Inteligencia Artificial, Mexico Itzamá López-Yáñez,Laboratorio de Inteligencia Artificial, Mexico
WP
100
Most models of Bidirectional associative memories intend to achieve that all trained pattern correspond to stable states; however, this has not been possible. Also, none of the former models has been able to recall all the trained patterns. In this work we introduce a new model of bidirectional associative memory which is not iterative and has no stability problems. It is based on the Alpha-Beta associative memories. This model allows perfect recall of all trained patterns, with no ambiguity and no conditions. An example of finger-print recognition is presented. WP-17 PID: 5191 A Soft Computing Based Approach for Modeling of Chaotic Time Series Jayashri Vajpai, J. N. V. University, India Arun JB, Government Polytechnic College Campus, India Nonlinear dynamic time series modeling is a generic problem, which permeates all fields of science. The authors have developed a soft computing based methodology for the modeling of systems represented by such series. The soft computing techniques that are a consortium of emerging technologies, have recently provided an alternative approach to mathematical modeling. The implementation of soft computing is based on the exploitation of the tolerance for imprecision, uncertainty and partial truth to achieve tractability, robustness and low cost solution. Fuzzy logic, neural networks and genetic algorithms are considered to be principal constituents of soft computing. Of these, the first component is primarily concerned with imprecision of data and information, the second with learning, and the third with optimization. In many applications, it is advantageous to exploit the synergism of these methods by using them in combination, rather than alone. The proposed model is based on Kasabov’s Evolving Fuzzy Neural Network and employs a genetic algorithm based method for the optimization of the most important parameters that govern the development of its structure. The well-examined Box Jenkins problem, to predict future values of the time series, based on the past history, is used as an illustrative example to demonstrate the potential of the proposed Genetic Evolving Fuzzy Neural Network (GEFuNN) model. The proposed methodology may find applications in the areas of signal processing, control, weather forecasting, economic and business planning and several other fields. WP-18 PID: 5198 A Novel Multiplier For Achieving the Programmability of Cellular Neural Network Peng Wang, Tsinghua University, China Xun Zhang, Tsinghua University, China Dongming Jin, Tsinghua University, China A novel CMOS four-quadrant analog-digital multiplier for implementing a programmable Cellular Neural Network (CNN) is presented. The circuit, which can be fabricated in a standard CMOS process, performs the four-quadrant weighting of interconnect signals. Using this multiplier a programmable CNN neuron can be implemented with little expense. Both simulation and test results are given for the circuit fabricated in a standard, mixed signal, 0.18µm, CMOS process. According to this design, one input is analog voltage and the other input is digital signal. The linearity deviation is less than 1% in the dynamic range (1.0V, 2.2V) centered on Vref=1.6V. The power supply voltage is 3.3V.
WP
101
WP-19 PID: 5202 Genetic Algorithm Approach to Image Reconstruction in Electrical Impedance Tomography Ho-Chan Kim, Cheju National University, Korea Chang-Jin Boo, Cheju National University, Korea Min-Jae Kang, Cheju National University, Korea In electrical impedance tomography (EIT), various image reconstruction algorithms have been used in order to compute the internal resistivity distribution of the unknown object with its electric potential data at the boundary. Mathematically the EIT image reconstruction algorithm is a nonlinear ill-posed inverse problem. This paper presents a genetic algorithm technique for the solution of the static EIT inverse problem. The computer simulation for the 32 channels synthetic data shows that the spatial resolution of reconstructed images in the proposed scheme is improved compared to that of the modified Newton–Raphson algorithm at the expense of increased computational burden. WP-20 PID: 5267 Neural Networks Based Scalable Fast Intra Prediction Algorithm in H.264 Encoder Jung-Hee Suk, Kyungbook National University, Korea Jin-Seon Youn, Kyungbook National University, Korea Jun Rim Choi, Kyungbook National University, Korea In this paper, we propose a neural network-based scalable fast intra prediction algorithm in H.264 in order to reduce redundant calculation time by selecting the best mode of 4×4 and 16×16 intra prediction. In this reason, it is possible to encode compulsively by 4×4 intra prediction mode for current MB (macro block)’s best prediction mode without redundant mode decision calculation in accordance with neural network’s output resulted from co-relation of adjacent encoded four left, up-left, up and up-right blocks. If there is any one of MBs encoded by 16×16 intra prediction among four MBs adjacent to current MB, the probability of re-prediction into 16×16 intra prediction will become high. We can apply neural networks in order to decide whether to force into 4×4 intra prediction mode or not. We can also control both the bit rates and calculation time by modulating refresh factors and weights of neural network’s output depend on error back-propagation, which is called refreshing. In case of encoding several video sequences by the proposed algorithm, the total encoding time of 30 input I frames are reduced by 20% ~ 65% depending upon the test vector compared with JM 8.4 by using neural networks and by modulating scalable refreshing factor. On the other hand, total encoding bits are increased by 0.8% ~ 2.0% at the cost of reduced SNR of 0.01 dB. WP-21 PID: 5279 Towards Hardware Acceleration of Neuroevolution for Multimedia Processing Applications on Mobile Devices Daniel Larkin, Dublin City University, Ireland Andrew Kinane, Dublin City University, Ireland Noel O’Connor, Dublin City University, Ireland
WP
102
This paper addresses the problem of accelerating large artificial neural networks (ANN), whose topology and weights can evolve via the use of a genetic algorithm. The proposed digital hardware architecture is capable of processing any evolved network topology, whilst at the same time providing a good trade o between throughput, area and power consumption. The latter is vital for a longer battery life on mobile devices. The architecture uses multiple parallel arithmetic units in each processing element (PE). Memory partitioning and data caching are used to minimise the effects of PE pipeline stalling. A first order minimax polynomial approximation scheme, tuned via a genetic algorithm, is used for the activation function generator. Efficient arithmetic circuitry, which leverages modified Booth recoding, column compressors and carry save adders, is adopted throughout the design. WP-22 PID: 5285 A Competitive Co-Evolving Support Vector Clustering Sung-Hae Jun, Cheongju University, Korea Kyung-Whan Oh, Sogang University, Korea The goal of clustering is to cluster the objects into groups that are internally homogeneous and heterogeneous from group to group. Clustering is an important tool for diversely intelligent systems. So, many works have been researched in the machine learning algorithms. But, some problems are still shown in the clustering. One of them is to determine the optimal number of clusters. In K-means algorithm, the number of cluster K is determined by the art of researchers. Another problem is an over fitting of learning models. The majority of learning algorithms for clustering are not free from the problem. Therefore, we propose a competitive co-evolving support vector clustering. Using competitive co-evolutionary computing, we overcome the over fitting problem of support vector clustering which is a good learning model for clustering. The number of clusters is efficiently determined by our competitive co-evolving support vector clustering. To verify the improved performances of our research, we compare competitive co-evolving support vector clustering with established clustering methods using the data sets form UCI machine learning repository. WP-23 PID: 5291 Pattern Discovery from Time Series Using Growing Hierarchical Self-Organizing Map Shiyuan Liu, Huazhong University of Science and Technology, China Li Lu, Huazhong University of Science and Technology, China Guanglan Liao, Huazhong University of Science and Technology, China Jianping Xuan, Huazhong University of Science and Technology, China Pattern discovery from time series is an important task in many applications. The unsupervised self-organizing map (SOM) has been widely used in data mining as well as in time series knowledge discovery. However, the traditional SOM has two main limitations: the static architecture and the lacking ability for the representing of hierarchical relations of the data. To overcome these limitations the growing hierarchical self-organizing map (GHSOM) is used to analyze time series in this paper. The experiments conducted on several data sets confirm that the GHSOM can form an adaptive architecture, which grows in size and depth during its training process, thus to unfold the hierarchical structure of the analyzed time series data. It is expected that this
WP
103
method will be effective and efficient to implement and will provide a useful practical tool for pattern discovery from large time series databases. WP-24 PID: 5307 Probabilistic Approximation under Incomplete Information Systems Wenhai Li, Huazhong University of Science and Technology, China Yucai Feng, Huazhong University of Science and Technology, China Zehua Lv, Huazhong University of Science and Technology, China This paper, firstly, analyzes the two semantics of unavailable values in incomplete information systems. For a given object, by applying the probability estimation of the unavailable attributes deriving from the available attributes to the neighbor objects, the suited degree of each neighbor object to the given object is depicted, therefore, the suited precision guaranteed neighborhood space is obtained. We show how to shrink the rules search space via variable precision rough set model based on this space, and also, we prove the incredibility degree of decision class is guaranteed by the two-level thresholds. The rule reducing method is indicated to be available by our experiments. WP-25 PID: 5328 Mental Representation and Processing Involved in Comprehending Korean Regular and Irregular Verb Eojeols: an fMRI and Reaction Time Study Hyungwook Yim, Korea University, Korea Changsu Park, Korea University, Korea Heuiseok Lim, Kichun Nam, Hanshin University, Korea The purpose of this study is to investigate the cortical areas involved in comprehending Korean regular and irregular verb Eojeols. Eojeols is the specific spacing unit of a sentence which is bigger than a word but smaller than phrase. This study showed that there is a distinction between the process and representation of regularly and irregularly inflected verbs in Korean using neuroimaging and behavioral method. WP-26 PID: 5346 Prediction error of a fault tolerant neural network John Sum, Chung Shan Medical University, Taiwan Chi-sing Leung, City University of Hong Kong, Hong Kong Kevin Ho, Providence University, Taiwan For more than a decade, prediction error has been one powerful tool to measure the performance of a neural network. In this paper, we extend the technique to a kind of fault tolerant neural network. Consider a neural network to be suffering from multiple-node fault, a formulae similar to that of Generalized Prediction Error has been derived. Hence, the effective number of parameter of such a fault tolerant neural network is obtained. A difficulty in obtaining the mean prediction error is discussed and then a simple procedure for estimation of the prediction error empirically is suggested. WP-27 PID: 5360
WP
104
From Hopfield Nets to Pulsed Neural Networks Ana M. G. Guerreiro, Rio Grande do Norte Federal University, Brasil Carlos A. Paz de Araujo, University of Colorado, United States of America Considering the first two generations of Artificial Neural Networks, Hopfield model is the only active system. Studying this type of network, a relation between this artificial neural network and the third generation, characterized by spiking neurons, was noticed. This paper presents the relationship between the Hopfield Neural Networks and the Pulsed Neural Networks. This relation is shown by the integration of the Hopfield neuron model, ending in an integrate-and-fire model, with the appropriate choice of the input kernels. WP-28 PID: 5378 Forecasting the Flow of Data Packets for Website Traffic Analysis - ASVR-tuned ANFISCH Approach Bao Rong Chang, National Taitung University, Taiwan Shi-Huang Chen, Shu-Te University, Taiwan Hsiu Fen Tsai, Shu-Te University, Taiwan Forecast of the flow of data packets between client and server for a website traffic analysis is viewed as a part of web analytics. Thousands of websmart businesses depend on web analytics to improve website conversions, reduce marketing costs, website optimization, website monitoring and provide a higher level of service to their customers and partners. This paper particularly intends to develop a high-accuracy prediction approach as the need for a website traffic analysis. The proposed composite model (ASVR-ANFIS/NGARCH) is schemed to build a systematic structure such that it is not only to improve the predictive accuracy because of resolving the problems of the overshoot and volatility clustering simultaneously, but also to boost website tracking capacity helping each webmaster to optimize their website, maximize online marketing conversions and lead campaign tracking. WP-29 PID: 5411 Scaling of learning target firing rate patterns with modulated STDP Razvan V. Florian, Center for Cognitive and Neural Studies (Coneural), Romania Spike-timing dependent plasticity (STDP) is a mechanism observed in the brain that induces synaptic changes that depend on the relative timing of pre- and postsynaptic spikes. The modulation of STDP with a global reward signal has been shown in simulations to lead to reinforcement learning. Here we show that learning of firing rate patterns with modulated STDP scales well to networks of thousands of neurons, thus further supporting the biological relevance of this learning rule. For the studied setup, learning time is several tens of seconds and does not depend on network size. WP-30 PID: 5425 An Electromechanical Neural Network Robotic Model of the Human Body and Brain. Sensory-Motor Control by Reverse Engineering Biological Somatic Sensors. Alan Rosen, Machine Cosciousness Inc., USA David B. Rosen, Machine Cosciousness Inc., USA
WP
105
This paper presents an electromechanical robotic model of the human body and brain. The model is designed to reverse engineer some biological functional aspects of the human body and brain. The functional aspects includes reverse engineering, a) biological perception by means of “sensory monitoring” of the external world, b) “self awareness” by means of monitoring the location and identification of all parts of the robotic body, and c) “biological sensory motor control” by means of feedback monitoring of the internal reaction of the robotic body to external forces. The model consists of a mechanical robot body controlled by a neural network based controller. WP-31 PID: 5428 Application of Self-Organizing Map (SOM) for Cerebral Cortex Reconstruction Cheng-Hung Chuang, Academia Sinica, Taiwan Philip E. Cheng, Academia Sinica, Taiwan Michelle Liou, Academia Sinica, Taiwan Cheng-Yuan Liou, National Taiwan Univ., Taipei, Taiwan Yen-Ting Kuo, National Taiwan Univ., Taipei, Taiwan This paper presents the application of a self-organizing map (SOM) model for the reconstruction of cerebral cortex from MRI images. The cerebral cortex is an important tissue for many brain science or medicine related researches. Since it is difficult to extract the highly folded and buried cortical surface, we apply the SOM model to deform the easily extracted white matter surface on a layered distance map to obtain the cortical surface. The layered distance map is calculated according to the extracted white matter surface. The proposed method can extract the proper cortical surface and make the measure of the cortical thickness easy. The simulations on a simple artificial image and T1-weighted MRI images show that the proposed algorithm is robust to reconstruct cerebral cortex. WP-32 PID: 5551 Multistage Blind Source Separation and Deconvolution for Convolutive Mixture of Speech Signals Yanxue Liang, Tokyo Institute of Technology, Japan Fengyu Cong, Noise, Shanghai Jiao Tong University, China Ichiro Hagiwara, Tokyo Institute of Technology, Japan In present paper, a new MBD based multistage method is proposed to solve well known whitening effect, in which contributions of each source to every microphone can be retrieved based on compensation matrix. In detail, MBD is first implemented, then compensation matrix is constructed, based on which contributions of sources to every microphone are retrieved, after that, to recover original signals perfectly, remainder work is Single Input and Multi- Output (SIMO) dereverberation that can be easily carried out. On the other hand, time domain MBD algorithm is difficult to converge without good initialization. To resolve this problem, Null Beamforming (NBF) combined with FastICA based Direction of Arrival (DOA) estimation is proposed. Such initialization generally guarantees convergence of time domain MBD so that compensation matrix can be constructed stably. Finally, experiment demonstrates validity and superiority of our scheme over other methods.
WP
106
WP-33 PID: 5565 A Hardware Implementation of Wavelet Neural Network in FPGAs Ali Karabiyik, Ege University, Turkey Aydogan Savran, Ege University, Turkey In this paper, hardware implementation of a wavelet neural network (WNN) is described. The WNN is developed in MATLAB and implemented on a Field-Programmable Gate Array (FPGA) device. The structure of the WNN is similar to the radial basis function (RBF) network, except that here the radial basis functions are replaced by orthonormal scaling functions. The training of the WNN is simplified due to the orthonormal properties of the scaling functions. The performances of the proposed WNN are tested by applying for the function approximation, system identification and the classification problems. Because of their parallel processing properties, the FPGAs provide good alternative in real-time applications of the WNN. By means of the simple scaling function used in the WNN architecture, it can be favorable to multilayer feedforward neural network and the RBF Networks implemented on the FPGA devices.
TA401
107
Thursday Session A, 5 October 2006
TA401 Invited talk by Shiro Usui Invited talk by Shoujue Wang Section Chair: Laiwan CHAN, The Chinese University of Hong Kong
TA402
108
TA402 Special Session Neurodynamics Section Chair: Rubin WANG, East China University of Science and Technology TA402-1 PID: 6001 Functional Differences between the Spatio-Temporal Learning Rule (STLR) and Hebb type (HEBB) in single pyramidal cells in the hippocampal CA1 Area JMinoru Tsukada, Tamagawa University, Japan Yoshiyuki Yamazaki, Tamagawa University, Japan The spatio-temporal learning rule (STLR), proposed as a non-Hebb type by Tsukada et al. (1996 [1], 2005 [2]), consists of two distinctive factors; cooperative plasticity without a postsynaptic spike, and its temporal summation. On the other hand, Hebb (1949 [3]) proposed the idea (HEBB) that synaptic modification is strengthened only if the pre- and post-synaptic elements are activated simultaneously. We have shown, experimentally, that both STLR and HEBB coexist in single pyramidal cells of the hippocampal CA1 area. The functional differences between STLR and HEBB in dendrite (local)- soma (global) interactions in single pyramidal cells of CA1 and the possibility of reinforcement learning were discussed. TA402-2 PID: 5554 Anti-phase of the leakage current and the sum of the other currents in membrane models J X Xu, Xi’an Jiaotong University, China; Z F Yue, Xi’an Jiaotong University, China; X C Yao, Xi’an Jiaotong University, China The responses of several conductance-based neural membrane models to sustain and chaos stimulus are carried out through numerical method. The Lorenz model is adapted as the chaos stimulation. It is shown that the sum of sodium and potassium current items and the leakage current in the Hodgkin-Huxley models are anti-phase. That is to say when the sum arrives at the minimum value, the leakage arrives at the maximum value. And the sum of potassium and the calcium current items and the leakage in the Morris-Lecar model are anti-phase too. This character is ubiquitous in the Hodgkin-Huxley model, the Morris-Lecar model and many other conductance-based models. But, as we know, the mechanism under which the relation produces is unknown. TA402-3 PID: 4840 Self-organizing Rhythmic Patterns with Spatio-temporal spikes in Class I and Class II Neural Networks Ryosuke HOSAKA, Saitama University, Japan; Tohru IKEGUCHI, Saitama University, Japan; Kazuyuki AIHARA, University of Tokyo, Japan Regularly spiking neurons are classified into two categories, Class I and Class II, by their firing properties for constant inputs. To investigate how the firing properties of single neurons affect to ensemble rhythmic activities in neural networks, we constructed different types of neural networks whose excitatory neurons are the Class I neurons or the
TA402
109
Class II neurons. The networks were driven by random inputs and developed with STDP learning. As a result, the Class I and the Class II neural networks generate different types of rhythmic activities: the Class I neural network generates slow rhythmic activities, and the Class II neural network generates fast rhythmic activities. TA402-4 PID: 5502 Fatigue-Induced Asymmetric Hemispheric Plasticity: An Event Related Potentials Study Ling-Fu Meng, Chang Gung University, Taiwan; Chiu-Ping Lu, Chang Gung University, Taiwan; Bo-Wei Chen, Chang Gung University, Taiwan; Ching-Horng Chen, Chang Gung University, Taiwan Based on a preliminary case study, we conducted an event related potentials (ERPs) research to explore the relationship between repetitive finger tapping and brain electrophysiological potentials. This present study found that the errors increased with motor repetitions during tapping tasks by right hand especially in the third stage. We defined this stage as the fatigue stage and the first stage as the initial stage. In the fatigue stage, the decreased N1 amplitudes (30-80 ms) with the right fronto-central and right central electrodes (FC4 and C4) were observed, while comparing with the initial stage. Moreover, the pronounced P2 amplitude (150-200 ms) and increased signal with time on right hemisphere (F4 and C4 electrodes) under fatigue state were noticed. Conversely, the contralateral left electrodes (FC3, C3, and F3) did not show aforementioned N1 and P2 differences between two stages. After using the Frequency Extraction method, a clear lateralized pattern in the fatigue stage was found. The left hemisphere showed lower and the right hemisphere showed higher alpha frequency phase content evolution. It was concluded that fatigue did lower the involvement of some areas in the brain but also did make right hemisphere take on more workload during the tapping task with right hand. We call this compensatory change as fatigue-induced asymmetric hemispheric plasticity. Besides, less signal change between two hemispheres in the fatigue stage was also found. Therefore, the mechanism of transcallosal interaction is strongly related to the fatigue state induced by the motor repetitions. TA402-5 PID: 5044 Time Variant Causality Model Applied in Brain Connectivity Network Based on Event Related Potential Kai Yin, Beijing Normal University, China; Xiao-Jie Zhao, Beijing Normal University, China; Li Yao, Beijing Normal University, China; Granger causality model mostly used to find the interaction between different time series are more and more applied to natural neural network at present. Brain connectivity network that could imply interaction and coordination between different brain regions is a focused research of brain function. Usually synchronization and correlation are used to reveal the connectivity network based on event-related potential (ERP) signals. However, these methods lack the further information such as direction of the connectivity network. In this paper, we performed an approach to detect the direction by Granger causality model. Considering the non-stationary of ERP data, we used traditional recursive least square (RLS) algorithm to calculate time variant Granger causality. In particular, we extended the method on the
TA402
110
significance of causality measures in order to make results more reasonable. These approaches were applied to the classic Stroop cognitive experiment to establish the causality network related to attention process mechanism.
TA403
111
TA403 SVM II Section Chair: James KWOK, Hong Kong University of Science and Technology TA403-1 PID: 5471 Support Vector Machines for Imbalanced Training Sets Jia Yinshan, Liaoning University of Petroleum and Chemical Technology, China; Wang Yumei, Qingdao Port Co, Ltd., China; ?-Support Vector Machine is one of the most widely used support vector machines because it provides a way to control the fraction of margin errors and the fraction of support vectors. However, being biased towards the class with the more training samples prevents it from being applied to some applications in which the size of classes is uneven. Although some updated ?-SVMs have been proposed to solve this problem, there are some issues in the formulations of these updated ?-SVMs. In this paper, a new ?-SVM, Double-? Support Vector Machine, is proposed. I t introduces ?+1 and ?-1 to control the upper bound on the fraction of bound support vectors and the lower bound on the fraction of support vectors of the positive class and the negative class respectively. Moreover, it reduces the complexity of formulations by eliminating a redundant constraint in ?-SVM formulations. TA403-2 PID: 5004 A Novel Sequential Minimal Optimization Algorithm for Support Vector Regression Jun Guo, Kyushu University, Japan; Norikazu Takahashi, Kyushu University, Japan; Tetsuo Nishi, Waseda University, Japan; A novel sequential minimal optimization (SMO) algorithm for support vector regression is proposed. This algorithm is based on Flake and Lawrence’s SMO in which convex optimization problems with l variables are solved instead of standard quadratic programming problems with 2l variables where l is the number of training samples, but the strategy for working set selection is quite different. Experimental results show that the proposed algorithm is much faster than Flake and Lawrence’s SMO and comparable to the fastest conventional SMO. TA403-3 PID: 5223 EUS SVMs: Ensemble of Under-Sampled SVMs for Data Imbalance Problems Pilsung Kang, Seoul National University, Korea; Sungzoon Cho, Seoul National University, Korea; Data imbalance occurs when the number of patterns from a class is much larger than that from the other class. It often degenerates the classification performance. In this paper, we propose an Ensemble of Under-Sampled SVMs or EUS SVMs. We applied the proposed method to two synthetic and six real data sets and we found that it outperformed other methods, especially when the number of patterns belonging to the minority class is very small.
TA403
112
TA403-4 PID: 5286 Depth-First Spanning 2-class SVM Tree Shaoning Pang, Auckland University of Technology, New Zealand; Nikola Kasabov, Auckland University of Technology, New Zealand; This article proposed a depth-first spanning SVM tree for the classification of 2-class pattern (2-SVMT). The proposed 2-SVMT simulates a common human-intelligence on 2-class classification with a recursive scalable data partition and 2-class SVM classification, where a set of 2-class SVMs are aggregated in a depth-first spanning ordered hierarchy. The proposed 2-SVMT method is able to cope with effectively both class imbalance and class overlap simply by combining a family of concurrent SVMs, while does not utilize adjustments of decision probabilities and SVM learning function, nor undersampling, oversampling or resampling. TA403-5 PID: 5062 Filtering E-mail Based on Fuzzy Support Vector Machines and Aggregation Operator Jilin Yang, Xihua University, China; Hong Peng, Xihua University, China; Zheng Pei, Xihua University, China; How to filter emails is a problem for Internet users. Support vector machine (SVM) is a valid filtering emails method. As it is well known, there exists uncertainty in deciding the legitimate email by Internet users. To formalize the uncertainty, the legitimate email is understood as fuzzy concept on a set of email samples in this paper, its membership function is obtained by aggregating opinions of Internet users, and aggregation operator is ordered weighted averaging (OWA) operator. Due to email training samples with membership degrees of the legitimate email, fuzzy support vector machine (FSVM) is adopted to classify emails, and penalty factor of FSVM is decided by content-specific misclassification costs. The advantages of our method are: 1) uncertainty of the legitimate email, i.e., membership degree, is considered in classifying emails, and a method to obtain membership degree is given; 2) content-specific misclassification costs is used to decide penalty factor of FSVM. Simulative experiments are shown to the effectiveness and human consistent of our method.
TA404
113
TA404 Image Processing II Section Chair: Jagath RAJAPAKSE, Nanyang Technological University TA404-1 PID: 4572 Fast and adaptive low-pass whitening filters for natural images Ling-Zhi Liao, Beijing Jiaotong University, China; Si-Wei Luo, Beijing Jiaotong University, China; Mei Tian, Beijing Jiaotong University, China; Lian-Wei Zhao, Beijing Jiaotong University, China A fast and simple solution was suggested to reduce the inter-pixels correlations in natural images, of which the power spectra roughly fell off with the increasing spatial frequency f according to a power law; but the 1/f exponent, was different from image to image. The essential of the proposed method was to flatten the decreasing power spectrum of each image by using an adaptive low-pass and whitening filter. The act of low-pass filtering was just to reduce the effects of noise usually took place in the high frequencies. The act of whitening filtering was a special processing, which was to attenuate the low frequencies and boost the high frequencies so as to yield a roughly flat power spectrum across all spatial frequencies. The suggested method was computationally more economical than the geometric covariance matrix based PCA method. Meanwhile, the performance degradations accompanied with the computational economy improvement were fairly insignificant. TA404-2 PID: 5470 Power spectral density based activation detection in functional MRI time-series Arun Kumar, Singapore Polytechnic, Singapore; Jagath C. Rajapakse, Engineering, Nanyang Technological University, Singapore; An efficient method for detecting activation on single and multiple cycle functional MRI (fMRI) data based on a hidden Markov model using power spectral density vectors is presented. Conventional methods of analysis of fMRI data are generally based on time-domain correlation analysis and concentrate mainly on the multiple epoch data and generally do not provide good results for single epoch data. The motivation to study and analyze single epoch data stems from the fact that certain experiments such as pain response, sleep onset, or administration of pharmacological agents, can only have a single or few trials. Our method has other advantages that it obviates the need to exclusively model the hemodynamic response function and also the voxels with delayed activation are correctly identified. We demonstrate the efficacy of our method to detect brain activations by using both synthetic and real fMRI data. TA404-3 PID: 4825 A Fast Directed Tree Based Neighborhood Clustering Algorithm for Image Segmentation Jundi Ding, Nanjing University of Aeronautics and Astronautics, China; SongCan Chen, Nanjing University of Aeronautics and Astronautics, China; RuNing Ma, Nanjing University of Aeronautics and Astronautics, China; Bo Wang, Nanjing University of Aeronautics and Astronautics, China
TA404
114
First, a modified Neighborhood-Based Clustering (MNBC) algorithm using the directed tree for data clustering is presented. It represents a dataset as some directed trees corresponding to meaningful clusters. Governed by Neighborhood-based Density Factor (NDF), it also can discover clusters of arbitrary shape and different densities like NBC. Moreover, it greatly simplifies NBC. However, a failure applying in image segmentation is due to an unsuitable use of Euclidean distance between image pixels. Second, Gray NDF (GNDF) is introduced to make MNBC suitable for image segmentation. The dataset to be segmented is all grays and thus MNBC has the constant computational complexity O (256). The experiments on synthetic datasets and real-world images show that MNBC outperforms some existing graph-theoretical approaches in terms of computation time as well as segmentation effect. TA404-4 PID: 5135 Fold Recognition using One-Class SVM Alexander Senf, The University of Kansas, USA Xue-wen Chen, The University of Kansas, USA Anne Zhang, The University of Kansas, USA The best protein structure prediction results today are achieved by incorporating initial structural prediction using alignments to known protein structures. The performance of these algorithms directly depends on the quality and significance of the alignment results. Support Vector Machines (SVMs) have shown great potential in providing good alignment results in cases where very low similarities to known proteins exist. In this paper we propose the use of a one-class SVM to reduce the computational resources required to perform SVM learning and classification. Experimental results show its efficiency compared to two-class SVM algorithms while producing results of similar accuracy. TA404-5 PID: 4828 Cascade of Selection and Fusion for Adaptive Classifier Mi Young Nam, Inha University, Korea; Eun Sung Jung, Inha University, Korea; Battulga Bayarsaikhan, Inha University, Korea; Suman Sedai, Inha University, Korea; Phill Kyu, Inha University, Korea; This paper proposes a novel adaptive classifier combination scheme based on the cascade of classifier selection and fusion, called adaptive classifier combination scheme (ACCS). In the proposed scheme, system working environment is learned and the environmental context is identified. Proposed t-test is used to search most effective classifier systems for each identified environmental context. The group of selected classifiers is combined based on t-test decision model for reliable fusion. The knowledge of individual context and its associated chromosomes representing the optimal classifier combination is stored in the context knowledge base. Once the context knowledge is accumulated the system can react to dynamic environment in real time. The proposed scheme has been tested in area of face recognition using standard FERET database, taking illumination as an environmental context. Experimental result showed that using context awareness in classifier combination provides robustness to varying environmental conditions.
TA405
115
TA405 Optimization Section Chair: Xiaolin HU, The Chinese University of Hong Kong TA405-1 PID: 5025 Solving Extended Linear Programming Problems Using a Class of Recurrent Neural Networks Xiaolin Hu, The Chinese University of Hong Kong, Hong Kong; Jun Wang, The Chinese University of Hong Kong, Hong Kong Extended linear programming (ELP) is an extension of classic linear programming in which the decision vector varies within a set. In previous studies in the neural network community, such a set is often assumed to be a box set. In the paper, the ELP problem with a general polyhedral set is investigated, and three recurrent neural networks are proposed for solving the problem with different types of constraints classified by the presence of bound constraints and equality constraints. The neural networks are proved stable in the Lyapunov sense and globally convergent to the solution set of corresponding ELP problems. Numerical simulations are provided to demonstrate the results. TA405-2 PID: 5055 A Recurrent Neural Network for Solving Non-smooth Convex Programming Problems Qingshan Liu, The Chinese University of Hong Kong, Hong Kong; Jun Wang, The Chinese University of Hong Kong, Hong Kong In this paper, a recurrent neural network model is proposed for solving non-smooth convex programming problems, which is a natural extension of the previous neural networks. By using the non-smooth analysis and the theory of differential inclusions, the global convergence of the equilibrium is analyzed and proved. One simulation example shows the convergence of the presented neural network. TA405-3 PID: 5110 A Swarm Optimization Model for Energy Minimization Problem of Early Vision Wenhui Zhou, ZheJiang University, China Lili Lin, ZheJiang Gongshang University, China Weikang Gu, ZheJiang University, China This paper proposes a swarm optimization model for energy minimization problem of early vision, which is based on a multicolony ant scheme. Swarm optimization is a new artificial intelligence field, which has been proved suitable to solve various combinatorial optimization problems. Compared with general optimization problems, energy minimization of early vision has its unique characteristics, such as higher dimensions, more complicate structure of solution space, and dynamic constrain conditions. In this paper, the vision energy functions are optimized by repeatedly minimizing a certain number of sub-problems according to divide-and-conquer principle, and each colony is allocated to optimize one sub-problem independently. Then an appropriate information exchange strategy between neighboring colonies and an adaptive method for dynamic problem are applied to implement global optimization. As a typical example, stereo
TA405
116
correspondence will be solved using the proposed swarm optimization model. Experiments show this method can achieve good results. TA405-4 PID: 5197 A Genetic-Inspired Multicast Routing Optimization Algorithm with Bandwidth and End-to-end Delay Constraints Sanghoun Oh, Gwangju Institute of Science and Technology, Korea; ChangWook Ahn, Samsung Advanced Institute of Technology, Korea; R.S. Ramakrishna, Gwangju Institute of Science and Technology, Korea This paper presents a genetic-inspired multicast routing algorithm with Quality of Service (i.e., bandwidth and end-to-end delay) constraints. The aim is to efficiently discover a minimum-cost multicast tree (a set of paths) that satisfactorily helps various services from a designated source to multiple destinations. To achieve this goal, state of the art genetic-based optimization techniques are employed. Each chromosome is represented as a tree structure of Genetic Programming. A fitness function that returns a tree cost has been suggested. New variation operators (i.e., crossover and mutation) are designed in this regard. Crossover exchanges partial chromosomes (i.e., sub-trees) in a positionally independent manner. Mutation introduces (in part) a new sub-tree with low probability. Moreover, all the infeasible chromosomes are treated with a simple repair function. The synergy achieved by combing new ingredients (i.e., representation, crossover, and mutation) offers an effective search capability that results in improved quality of solution and enhanced rate of convergence. Experimental results show that the proposed GA achieves minimal spanning tree, fast convergence speed, and high reliability. Further, its performance is better than that of a comparative reference.
TA406
117
TA406 Robotics and Control II Section Chair: Jih-Gau JUANG, National Taiwan Ocean University TA406-1 PID: 5249 Hybrid Intelligent PID Control for MIMO System Jih-Gau Juang, National Taiwan Ocean University, Taiwan; Kai-Ti Tu, National Taiwan Ocean University, Taiwan; Wen-Kai Liu, National Taiwan Ocean University, Taiwan This paper presents a new approach using switching grey prediction PID controller to an experimental propeller setup which is called the twin rotor multi-input multi-output system (TRMS). The goal of this study is to stabilize the TRMS in significant cross coupling condition and to experiment with set-point control and trajectory tracking. The proposed scheme enhances the grey prediction method of difference equation, which is a single variable second order grey model (DGM(2,1) model). It is performed by real-value genetic algorithm (RGA) with system performance index as fitness function. We apply the integral of time multiplied by the square error criterion (ITSE) to form a suitable fitness function in RGA. Simulation results show that the proposed design can successfully adapt system nonlinearity and complex coupling condition. TA406-2 PID: 4291 Reliable Robust Controller Design for Nonlinear State-delayed Systems Based on Neural Networks Yanjun Shen, Three Gorges University, China; Hui Yu, Three Gorges University, China; Jigui Jian, Three Gorges University, China An approach is investigated for the adaptive guaranteed cost control design for a class of nonlinear state-delayed systems. The nonlinear term is approximated by a linearly parameterized neural network (LPNN). A linear state feedback H1 control law is presented. An adaptive weight adjustment mechanism for the neural networks is developed to ensure H1 regulation performance. It is shown that the control gain matrices can be transformed into a standard linear matrix inequality problem and solved via a developed recurrent neural network. TA406-3 PID: 4163 Multi-Degree Prosthetic Hand Control Using a New BP Neural Network R. C. Wang, Tsinghua University, China; F. Li, Tsinghua University, China; M. Wu, Northwestern University, China; J. Z. Wang, Tsinghua University, China; L. Jiang, Harbin Institute of Technology, China; H. Liu, Harbin Institute of Technology, China A human-like multi-fingered prosthetic hand, HIT hand, has been developed in Harbin Institute of Technology. This paper presents a new pattern discrimination method for
TA406
118
HIT hand control. The method uses a bagged-BP neural network based on combing the BP neural networks using bagging algorithm. Bagging has been used to overcome the problem of limited number of training data in unimodal systems, by combining neural networks as weak learners. We compared the results of the bagging based BP network, using four features, with the results obtained separately from these uni-feature systems. The results show that the bagged-BP network improves both the accuracy and stability of the BP classifier. TA406-4 PID: 4808 Tracking control of a mobile robot with kinematic uncertainty using neural networks An-Min Zou, The Chinese Academy of Sciences, China; Zeng-Guang Hou, The Chinese Academy of Sciences, China; Min Tan, The Chinese Academy of Sciences, China; Xi-Jun Chen, The Chinese Academy of Sciences, China; Yun-Chu Zhang, The Chinese Academy of Sciences, Shandong University of Architecture and Engineering, China In this paper, a kinematic controller based on input-output linearization plus neural network (NN) controller is presented for tracking control of a mobile robot with kinematic uncertainty. The NN controller, whose parameters are tuned on-line, can deal with the uncertainty imposed on the kinematics model of mobile robots. The stability of the proposed approach is guaranteed by the Lyapunov theory. Simulation results show the efficiency of the proposed approach.
TA407
119
TA407 Evolutionary Algorithms II Section Chair: Myung Won KIM, Soongsil University TA407-1 PID: 5261 Induction of Optimized Fuzzy Decision Tree Using Genetic Algorithm Myung Won Kim, Soongsil University, Korea; Joung Woo Ryu, Intelligent Robot Research Division Electronics and Telecommunications Research Institute, Korea; Fuzzy rules are suitable for describing uncertain phenomena and natural for human understanding and they are, in general, efficient for classification. In addition, fuzzy rules allow us to effectively classify data having non-axis-parallel decision boundaries, which is difficult for the conventional attribute-based methods. In this paper, we propose an optimized fuzzy rule generation method for classification both in accuracy and comprehensibility (or rule complexity). We investigate the use of genetic algorithm to determine an optimal set of membership functions for quantitative data. In our method, for a given set of membership functions a fuzzy decision tree is constructed and its accuracy and rule complexity are evaluated, which are combined into the fitness function to be optimized. We have experimented our algorithm with several benchmark data sets. The experiment results show that our method is more efficient in performance and complexity of rules compared with the existing methods. TA407-2 PID: 4701 Integration of Genetic Algorithm and Cultural Algorithms for Constrained Optimization Fang Gao, Harbin Institute of Technology, China; Gang Cui, Harbin Institute of Technology, China; Hongwei Liu, Harbin Institute of Technology, China In this paper, we propose to integrate real coded genetic algorithm (GA) and cultural algorithms (CA) to develop a more efficient algorithm: cultural genetic algorithm (CGA). In this approach, GA?-4s selection and crossover operations are used in CA?-4s population space. GA?-4s mutation is replaced by CA based mutation operation which can attract individuals to move to the semi-feasible and feasible region of the optimization problem to avoid the ?-3eyeless?-4 searching in GA. Thus it is possible to enhance search ability and to reduce computational cost. This approach is applied to solve constrained optimization problems. An example is presented to demonstrate the effectiveness of the proposed approach. TA407-3 PID: 5204 UEAS: A Novel United Evolutionary Algorithm Scheme Fei Gao, Wuhan University of Technology, China Hengqing Tong, Wuhan University of Technology, China How to detect global optimums of the complex function is of vital importance in diverse scientific fields. Though stochastic optimization strategies simulating evolution process are proved to be valuable tools, the balance between exploitation and exploration of
TA407
120
which is difficult to be maintained. In this paper, some established techniques to improve the performance of evolutionary computation are discussed firstly, such as uniform design, deflection and stretching the objective function, and space contraction. Then a novel scheme of evolutionary algorithms is proposed to solving the optimization problems through adding evolution operations to the searching space contracted regularly with these techniques. A typical evolution algorithm differential evolution is chosen to exhibit the new scheme’s performance and the experiments done to minimize the benchmark nonlinear optimization problems and to detect nonlinear map’s unstable periodic points show the put approach is very robust. TA407-4 PID: 4459 Human Hierarchical Behavior Based Mobile Agent Control in ISpace with Distributed Network Sensors SangJoo Kim, Dong-eui Institute of Technology, Korea; TaeSeok Jin, DongSeo University, Korea; Hideki Hashimoto, the University of Tokyo, Japan; The aim of this paper is to investigate a control framework for mobile robots, operating in shared environment with humans. The Intelligent Space (iSpace) can sense the whole space and evaluate the situations in the space by distributing sensors. The mobile agents serve the inhabitants in the space utilizes the evaluated information by iSpace. The iSpace evaluates the situations in the space and learns the walking behavior of the inhabitants. The human intelligence manifests in the space as a behavior, as a response to the situation in the space. The iSpace learns the behavior and applies to mobile agent motion planning and control. This paper introduces the application of fuzzy-neural network to describe the obstacle avoidance behavior learned from humans. Simulation and experiment results are introduced to demonstrate the efficiency of this method. TA407-5 PID: 4469 Fuzzy-Neural-Network-Based Mobile Agent Control in the Intelligent Space which can Learn Human Behavior TaeSeok Jin, DongSeo University, Korea; YoungDae Son, DongSeo University, Korea; Hideki Hashimoto, the University of Tokyo, Japan; The knowledge of human walking behavior has primary importance for mobile agent in order to operate in the human shared space, with minimal disturb of other humans. This paper introduces such an observation and learning framework, which can acquire the human walking behavior from observation of human walking, using CCD cameras of the Intelligent Space. The proposed behavior learning framework applies Fuzzy-Neural Network (FNN) to approximate observed human behavior, with observation data clustering in order to extract important training data from observation. Preliminary experiment and results are shown to demonstrate the merit of the introduced behavior.
TA408
121
TA408 ICA/BSS I Section Chair: Kun ZHANG, The Chinese University of Hong Kong TA408-1 PID: 5521 K-hyperplanes clustering and its application to sparse component analysis Zhaoshui He, Brain Science Institute, RIKEN, Japan; Andrzej Cichocki, Brain Science Institute, RIKEN, Japan; Shengli Xie, South China University of Technology, China In this paper the K-hyperplanes clustering problem is discussed and we present a K-hyperplanes clustering algorithm, which can be applied to sparse component analysis (SCA) for linear model = + X AS V , where X is a m by T matrix of observation, A is an unknown m by n basis matrix and S is an unknown n by T matrix of sparse sources. The proposed algorithm is suitable for a relaxed case when more than one source signal achieves significant value at any time instant. More precisely, in this paper we propose a new algorithm which is suitable for the case when the (1 m-) source signals are simultaneously nonzero for sufficient number of samples, where m is the number of observation. In contrast to the conventional SCA algorithm which is based on the assumption that for each time, there is only one dominant component and others components are not significant. We assume that the sources can be only moderately sparse. However, the complexity of the algorithm is higher than that of the conventional SCA algorithms. We confirmed the validity and good performance of the proposed algorithm by computer simulation. TA408-2 PID: 5563 Newton-like Methods for Nonparametric Independent Component Analysis Hao Shen, National ICT Australia, Australia
�Knut H uper, National ICT Australia, Australia Alexander J. Smola, National ICT Australia, Australia The performance of ICA algorithms significantly depends on the choice of the contrast function and the optimisation algorithm used in obtaining the demixing matrix. In this paper we focus on the standard linear nonparametric ICA problem from an optimisation point of view. It is well known that after a pre-whitening process, the problem can be solved via an optimisation approach on a suitable manifold. We propose an approximate Newton’s method on the unit sphere to solve the one-unit linear nonparametric ICA problem. The local convergence properties are discussed. The performance of the proposed algorithms is investigated by numerical experiments. TA408-3 PID: 5243 A Hybrid Blind Signal Separation Algorithm: Particle Swarm Optimization on Feed-forward Neural Network Chan-Cheng Liu, National Dong Hwa University, Taiwan Tsung-Ying Sun, National Dong Hwa University, Taiwan Sheng-Ta Hsieh, National Dong Hwa University, Taiwan Chun-Ling Lin, National Dong Hwa University, Taiwan
TA408
122
Kan-Yuan Lee, National Dong Hwa University, Taiwan The blind signal separation problem (BSS) which involved linear mixing model and stationary source signals is focused in this paper. In the past, the neural network (NN) model is the popular architecture for separation, but its performance depends on initiation of weight strongly. In order to improve this problem to enhance global convergent, the genetic algorithm (GA) has been introduced for optimizing the weights of NN system recently. This paper, a novel evolution algorithm, particle swarm optimization (PSO) is introduced to optimize NN weights by us. Further, in simulation experiments of BSS, it is demonstrated that the PSO-based NN system has better performance in terms of global searching, computational time, accuracy and efficiency than the GA-based NN system. TA408-4 PID: 4953 Electrogastrogram Extraction Using Independent Component Analysis with References Peng Cheng, Tsinghua University, China; Qian Xiang, Tsinghua University, China; Ye Datian, Tsinghua University, China Electrogastrogram (EGG) is a cutaneous measurement of gastric myoelectrical activity, which is usually covered by strong artifacts. This paper presents a new approach in the framework of constrained independent component analysis (ICA). The nonlinear uncorrelatedness between the desired component and references is introduced as a constraint. The results show that the proposed method can extract the desired component corresponding to gastric slow wave directly, avoiding the ordering indeterminacy in ICA. Furthermore, the perturbations in EGG can be suppressed effectively. In summary, it can be a useful method for EGG analysis in research and clinical practice. TA408-5 PID: 5326 Performance evaluation of directionally constrained filterbank ICA on blind source separation of noisy observations Chandra Shekhar Dhir, Korea Advanced Institute of Science and Technology, Korea; Hyung-Min Park, Carnegie Mellon University, USA; Soo-Young Lee, Korea Advanced Institute of Science and Technology, Korea Separation performance of directionally constrained filterbank ICA is evaluated in presence of noise with different spectral properties. Stationarity of mixing channels is exploited to introduce directional constraint on the adaptive sub-band separation networks of filterbank-based blind source separation approach. Directional constraints on demixing network improve separation of source signals from noisy convolved mixtures, when significant spectral overlap exists between the noise and the convolved mixtures. Observations corrupted with low frequency noises exhibit slight improvement in the separation performance as there is less spectral overlap. Initialization and constraining of sub-band demixing network in accordance to the spatial location of source signals result in faster convergence and effective permutation correction, irrespective, of the nature of additive noise.
TB401
123
Thursday Session B, 5 October 2006
TB401 Invited talk by Soo-Young Lee Invited talk by Nik Kasabov Section Chair: Shiro USUI, RIKEN
TB402
124
TB402 Cognitive Processing / Learning III Section Chair: Minho LEE, Kyungpook National University TB402-1 PID: 5349 Sequence Disambiguation by Functionally Divided Hippocampal CA3 Model Toshikazu Samura, Keio University, Japan; Motonobu Hattori, University of Yamanashi, Japan; Shun Ishizaki, Keio University, Japan Many sequences are acquired in the hippocampus. Sequences are represented by a series of events which are associative representations of information and they are linked with each other by common representations. The sequences must, however, be recalled individually, even if the sequences are overlapped at some representations: otherwise we couldn’t retrieve own precious memory correctly. Therefore, sequence disambiguation is an essential function to separate the overlapped sequences. In this study, we focus on the location dependencies of the hippocampal CA3 region which are obtained from very detailed anatomical findings of the CA3 and the physiological findings of Spike-Timing Dependent Plasticity (STDP), and suggest that the CA3 is functionally divided by its location dependencies. Moreover, we show that functionally divided CA3 could generate a code for sequence disambiguation by using computer simulations. Finally, we also suggest that the sequence disambiguation can be realized between the CA3 and the hippocampal CA1 region. TB402-2 PID: 5516 Learning with Incrementality Abdelhamid Bouchachia, University of Klagenfurt, Austria Learning with adaptivity is a key issue in many nowadays applications. The most important aspect of such an issue is incremental learning (IL). This latter seeks to equip learning algorithms with the ability to deal with data arriving over long periods of time. Once used during the learning process, old data is never used in subsequent learning stages. This paper suggests a new IL algorithm which generates categories. Each is associated with one class. To show the efficiency of the algorithm, several experiments are carried out. TB402-3 PID: 4669 A Neuropsychologically-Inspired Computational Approach to The Generalization of Cerebellar Learning S. D. Teddy, Nanyang Technological University, Singapore; E. M-K. Lai, Nanyang Technological University, Singapore; C. Quek, Nanyang Technological University, Singapore The CMAC neural network is a well-established computational model of the human cerebellum. A major advantage is its localized generalization property which allows for efficient computations. However, there are also two major problems associated with this localized associative property. Firstly, it is difficult to fully-train a CMAC network as the
TB402
125
training data has to fully cover the entire set of CMAC memory cells. Secondly, the untrained CMAC cells give rise to undesirable network output when presented with inputs that the network has not previously been trained for. To the best of the authors’ knowledge, these issues have not been sufficiently addressed. In this paper, we propose a neuropsychologically inspired computational approach to alleviate the above-mentioned problems. Motivated by psychological studies on human motor skill learning, a ?patching? algorithm is developed to construct a plausible memory surface for the untrained cells in the CMAC network. We demonstrate through the modeling of the human glucose metabolic process that the ?patching? of untrained cells offers a satisfactory solution to incomplete training in CMAC. TB402-4 PID: 5443 Lability of Reactivated Human Declarative Memory Fumiko Tanabe, Tokyo Institute of Technology, Sony Computer Science Laboratories, Japan; Ken Mogi, Sony Computer Science Laboratories, Tokyo Institute of Technology, Japan Memory consolidation is an increasingly important cortical process in which a new memory is transferred to a stable state over time. Memory is said to be labile when certain cognitive and/or pharmacological processes lead to its partial or total destruction. Classical theory held that once consolidated, memory is not susceptible to disruption. However, recent experiments have suggested that when a consolidated memory is reactivated, it can become labile again (Nader 2003). A process of reconsolidation is likely to be required for the activated memory (reconsolidation hypothesis). Here we investigate the lability of human declarative memory. The results show that a reactivated declarative memory becomes labile under certain conditions. It is suggested that declarative memory in an active recollected state becomes susceptible to modification, editing, and even erasure in some extreme cases. TB402-5 PID: 4609 Functional Connectivity in the Resting Brain: an Analysis Based on ICA Xia Wu, Beijing Normal University, China; Li Yao, Beijing Normal University, China; Zhi-ying Long, University of California at San Diego, USA; Jie Lu, Xuan Wu Hospital of Beijing, China; Kun-cheng Li, Xuan Wu Hospital of Beijing, China The functional connectivity of the resting state, or default mode, of the human brain has been a research focus, because it is reportedly altered in many neurological and psychiatric disorders. Among the methods to assess the functional connectivity of the resting brain, independent component analysis (ICA) has been very useful. But how to choose the optimal number of separated components and the best-fit component of default mode network are still problems left. In this paper, we used three different numbers of independent components to separate the fMRI data of resting brain and three criterions to choose the best-fit component. Furthermore, we proposed a new approach to get the best-fit component. The result of the new approach is consistent with the default-mode network.
TB403
126
TB403 Special Session -Extensions of Independent Component Analysis II Section Chair: Kun ZHANG, The Chinese University of Hong Kong TB403-1 PID: 4915 Multilayer Nonnegative Matrix Factorization Using Projected Gradient Approaches Andrzej CICHOCKI, Laboratory for Advanced Brain Signal Processing RIKEN BSI, Japan; Rafal ZDUNEK, Laboratory for Advanced Brain Signal Processing RIKEN BSI, Japan The most popular algorithms for Nonnegative Matrix Factorization (NMF) belong to the class of multiplicative Lee-Seung algorithms which have usually relative low complexity but are characterized by slow-convergence and risk to stuck in local minima. In this paper, we discuss the additive algorithms based on three different variations of a projected gradient approach. Additionally, we shortly discuss a novel multi-layer approach to NMF algorithms, which in general, consider- ably improves the performance of all NMF algorithms. We demonstrate that this approach (the multi-layer system with projected gradient algorithms) can usually give much better performance than standard multiplicative algorithms, especially, if data are ill-conditioned, badly-scaled, and/or a number of observations are only slightly greater than a number of sources. Our new implementations of NMF are demonstrated with the simulations performed for Blind Source Separation (BSS) data. TB403-2 PID: 5561 DSS: a flexible framework for source separation Jaakko S??¨¬arel??¨¬a, Helsinki University of Technology, Finland; Denoising source separation is a recently introduced framework for building source separation algorithms around denoising procedures. The denoising function codes the prior or acquired information that is needed to separate the source from the other sources and noise. The resulting algorithms are computationally efficient and are thus suitable for vast datasets. In this paper, we give a practical introduction to the framework and see its uses in neuroinformatics. TB403-3 PID: 5547 Constrained ICA Based Ballistocardiogram and Electro-oculogram Artifacts Removal from Visual Evoked Potential EEG Signals Measured inside MRI Tahir Rasheed, Kyung Hee University, Korea; Myung Ho In, Kyung Hee University, Korea; Young-Koo Lee, Kyung Hee University, Korea; Sungyoung Lee, Kyung Hee University, Korea; Soo Yeol Lee, Kyung Hee University, Korea; Tae-Seong Kim, Kyung Hee University, Korea In the simultaneous acquisition of EEG and fMRI, analysis of EEG signals is a difficult task due to ballistocardiogram (BCG) and electro-oculogram (EOG) artifacts. It gets worse if evoked potentials are measured inside MRI for their minute responses in
TB403
127
comparison to the spontaneous brain responses. In this paper, we propose a new method for removing both artifacts simultaneously from the evoked EEG signals acquired inside MRI using constrained Independent component analysis (cICA). With properly designed reference functions for the BCG and EOG artifacts as constraints, cICA identifies the independent components (ICs) corresponding to the artifacts. Then artifact-removed EEG signals are reconstructed after removing the identified ICs to obtain evoked potentials. To evaluate our proposed technique, we have removed the artifacts with cICA and the standard template subtraction technique and generated visual evoked potentials (VEPs) respectively which are compared to the VEPs obtained from EEG signals measured outside MRI. Our results indicate that our cICA technique performs better than the standard BCG artifact removal methods with some efficient features. TB403-4 PID: 5551 Multistage Blind Source Separation and Deconvolution for Convolutive Mixture of Speech Signals Yanxue Liang,Tokyo Institute of Technology, Japan; Fengyu Cong, Noise, Shanghai Jiao Tong University, China; Ichiro Hagiwara, Tokyo Institute of Technology, Japan In present paper, a new MBD based multistage method is proposed to solve well known whitening effect, in which contributions of each source to every microphone can be retrieved based on compensation matrix. In detail, MBD is first implemented, then compensation matrix is constructed, based on which contributions of sources to every microphone are retrieved, after that, to recover original signals perfectly, remainder work is Single Input and Multi- Output (SIMO) dereverberation that can be easily carried out. On the other hand, time domain MBD algorithm is difficult to converge without good initialization. To resolve this problem, Null Beamforming (NBF) combined with FastICA based Direction of Arrival (DOA) estimation is proposed. Such initialization generally guarantees convergence of time domain MBD so that compensation matrix can be constructed stably. Finally, experiment demonstrates validity and superiority of our scheme over other methods. TB403-5 PID: 5558 Extensions of ICA for Causality Discovery in the Hong Kong Stock Market Kun Zhang, The Chinese University of Hong Kong, Hong Kong; Lai-Wan Chan, The Chinese University of Hong Kong, Hong Kong; Recently independent component analysis (ICA) has been proposed for discovery of linear, non-Gaussian, and acyclic causal models (LiNGAM). As in practice the LiNGAM assumption usually does not exactly hold, in this paper we propose some methods to perform causality discovery even when LiNGAM is violated. The first method is ICA with a sparse separation matrix. By incorporating a suitable penalty term, the separation matrix produced by this method tends to satisfy the LiNGAM assumption. The other two methods are proposed to tackle nonlinearity in the data generation procedure, which violates the LiNGAM assumption. In the second method, the post-nonlinear mixing ICA model is exploited to do causality discovery when the nonlinearity is componentwise. The third method is proposed for the case where the nonlinear distortion in data generation is of arbitrary form, but smooth and weak. The separation system for such
TB403
128
data is a linear transformation coupled with a nonlinear one, and the nonlinear one is as weak as possible such that it can be neglected when performing causality discovery. The linear causal relations in the data are then revealed. The proposed methods are applied to discover the causal relations in the Hong Kong stock market, and the last method works very well. The resulting causal diagram shows some interesting information in the stock market.
TB404
129
TB404 Pattern Classification/Classifiers & Classification III (Face Analysis & Processing) Section Chair: Salim BOUZERDOUM, University of Wollongong TB404-1 PID: 4637 A Passport Recognition and Face Verification Using Enhanced Fuzzy Neural Network and PCA Algorithm Kwang-Baek Kim, Silla University, Korea; Sungshin Kim, Pusan National University, Korea In this paper, passport recognition and face verification methods which can automatically recognize passport codes and discriminate forgery passports to improve efficiency and systematic control of immigration management are proposed. Adjusting the slant is very important for recognition of characters and face verification since slanted passport images can bring various unwanted effects to the recognition of individual codes and faces. The angle adjustment can be conducted by using the slant of the straight and horizontal line that connects the center of thickness between left and right parts of the string. Extracting passport codes is done by Sobel operator, horizontal smearing, and 8-neighbornood contour tracking algorithm. The proposed RBF network is applied to the middle layer of RBF network by using the fuzzy logic connection operator and proposing the enhanced fuzzy ART algorithm that dynamically controls the vigilance parameter. After several tests using a forged passport and the passport with slanted images, the proposed method was proven to be effective in recognizing passport codes and verifying facial images. TB404-2 PID: 5019 A Weighted FMM Neural Network and Its Application to Face Detection Ho-Joon Kim, Handong University, Korea; Juho Lee, KAIST, Korea; Hyun-Seung Yang, KAIST, Korea In this paper, we introduce a modified fuzzy min-max (FMM) neural network model for pattern classification, and present a real-time face detection method using the proposed model. The learning process of the FMM model consists of three sub-processes: hyperbox creation, expansion and contraction processes. During the learning process, the feature distribution and frequency data are utilized to compensate the hyperbox distortion which may be caused by eliminating the overlapping area of hyperboxes in the contraction process. We present a multi-stage face detection method which is composed of two stages: feature extraction stage and classification stage. The feature extraction module employs a convolutional neural network (CNN) with a Gabor transform layer to extract successively larger features in a hierarchical set of layers. The proposed FMM model is used for the pattern classification stage. Moreover, the model is utilized to select effective feature sets for the skin-color filter of the system. TB404-3 PID: 5173 Fast Learning for Statistical Face Detection Zhi-Gang Fan, Shanghai Jiao Tong University, China;
TB404
130
Bao-Liang Lu, Shanghai Jiao Tong University, China In this paper, we propose a novel learning method for face detection using discriminative feature selection. The main deficiency of the boosting algorithm for face detection is its long training time. Through statistical learning theory, our discriminative feature selection method can make the training process for face detection much faster than the boosting algorithm without degrading the generalization performance. Being different from the boosting algorithm which works in an iterative learning way, our method can directly solve the learning problem of face detection. Our method is a novel ensemble learning method for combining multiple weak classifiers. The most discriminative component classifiers are selected for the ensemble. Our experiments show that the proposed discriminative feature selection method is more efficient than the boosting algorithm for face detection. TB404-4 PID: 4905 Gender Classification Using Neural Network S. L. Phung, University of Wollongong, Australia; A. Bouzerdoum, University of Wollongong, Australia We propose a novel neural network for classification of visual patterns. The new network, called pyramidal neural network or PyraNet, has a hierarchical structure with two types of processing layers, namely pyramidal layers and 1-D layers. The PyraNet is motivated by two concepts: the image pyramids and local receptive fields. In the new network, nonlinear 2-D are trained to perform both 2-D analysis and data reduction. In this paper, we present a fast training method for the PyraNet that is based on resilient back-propagation and weight decay, and apply the new network to classify gender from facial images.
TB405
131
TB405 App./Biomedical II Section Chair: Michael SMALL, Hong Kong Polytechnic University TB405-1 PID: 5426 Mining Neural Network based Feature Association Rule for Protein Interaction Prediction Jae-Hong Eom, Seoul National University, Korea; Byoung-Tak Zhang, Seoul National University, Korea Prediction of protein interactions is one of the central problems in post-genomic biology. In this paper, we present an association rule-based protein interaction prediction method. We adopted neural network to cluster protein interaction data, and used information theory based feature selection method to reduce protein feature dimension. After model training, feature association rules are generated to interaction prediction by decoding a set of learned weights of trained neural network and by mining association rules. For model training, an initial network model was constructed with public Yeast protein interaction data considering their functional categories, set of features, and interaction partners. The prediction performance was compared with traditional simple association rule mining method. The experimental results show that proposed method has about 96.1% interaction prediction accuracy compared to simple association mining approach which achieved about 91.4% accuracy. TB405-2 PID: 4793 An empirical analysis of under-sampling techniques to balance a protein structural class dataset Marcilio C. P. de Souto, Federal University of Rio Grande do Norte; Valnaide G. Bittencourt, Federal University of Rio Grande do Norte; Jose A. F. Costa, Federal University of Rio Grande do Norte There has been a great deal of research on learning from imbalanced datasets. Among the widely used methods proposed to solve such a problem, the most common are based either on under or over sampling of the original dataset. In this work, we evaluate several methods of under-sampling, such as Tomek Links, with the goal of improving the performance of the classifiers generated by different ML algorithms (decision trees, support vector machines, among others) applied to problem of determining the structural similarity of proteins. TB405-3 PID: 5318 DRFE: Dynamic Recursive Feature Elimination for Gene Identification based on Random Forest Ha-Nam Nguyen, Hankuk Aviation University, Korea; Syng-Yup Ohn, Hankuk Aviation University, Korea; Determining the relevant features is a combinatorial task in various fields of machine learning such as text mining, bioinformatics, pattern recognition, etc. Several scholars have developed various methods to extract the relevant features but no method is really
TB405
132
superior. Breiman proposed Random Forest to classify a pattern based on CART tree algorithm and his method turns out good results compared to other classifiers. Taking advantages of Random Forest and using wrapper approach which was first introduced by Kohavi et. al, we propose an algorithm named Dynamic Recursive Feature Elimination (DRFE) to find the optimal subset of features for reducing noise of the data and increasing the performance of classifiers. In our method, we use Random Forest as induced classifier and develop our own defined feature elimination function by adding extra terms to the feature scoring. We conducted experiments with two public datasets: Colon cancer and Leukemia cancer. The experimental results of the real world data showed that the proposed method has higher prediction rate compared to the baseline algorithm. The obtained results are comparable and sometimes have better performance than the widely used classification methods in the same literature of feature selection. TB405-4 PID: 4830 Choosing Real-Time Predictors For Ventricular Arrhythmias Detection B Ribeiro, University of Coimbra, Portugal; A Marques, University of Coimbra, Portugal; J Henriques, University of Coimbra, Portugal; M Antunes, Hospitais da Universidade de Coimbra, Portugal The risk of developing life-threatening ventricular arrhythmias in patients with a structural heart disease is higher with increased occurrence of premature ventricular complex (PVCs). Therefore the reliable detection of these arrhythmias constitutes a challenge for a cardiovascular diagnostic system. While early diagnosis is critical, the task of its automatic detection and classification becomes crucial. Therefore the underlying models should be efficient yet ensuring robustness. Although neural networks (NNs) have proved successful in this setting, we show that kernel-based learning algorithms have superior performance. In particular recently developed sparse Bayesian methods such as, Relevance Vector Machines (RVMs), present a parsimonious solution when compared with Support Vector Machines (SVMs) yet revealing a competitive accuracy property. This can lead to significant reduction in the computational complexity of the decision function, thereby making them more suitable for real-time applications.
TB406
133
TB406 Learning Algorithms II Section Chair: Sabri ARIK, Istanbul University TB406-1 PID: 4396 Mixture of Neural Networks: Some Experiments with the Multilayer Feedforward Architecture “Joaqu??n Torres-Sospedra, Universitat Jaume I, Spain; Carlos Hern?andez-Espinosa, Universitat Jaume I, Spain; Mercedes Fern?andez-Redondo, Universitat Jaume I, Spain “Joaqu??n Torres-Sospedra, Universitat Jaume I, Spain; Carlos Hern?andez-Espinosa, Universitat Jaume I, Spain; Mercedes Fern?andez-Redondo, Universitat Jaume I, Spain A Modular Multi-Net System consists of some networks which solve partially a problem. The original problem has been decomposed into sub-problems and each network focuses on solving a sub-problem. The Mixture of Neural Networks consists of some expert networks which solve the sub-problems and a gating network which weights the outputs of the expert networks. The expert networks and the gating network are trained all together in order to reduce the correlation among the networks and minimize the error of the system. In this paper we present the Mixture of Multilayer Feedforward (MixMF) a method based on MixNN which uses Multilayer Feedforward networks for the expert level. Finally, we have performed a comparison among Simple Ensemble, MixNN and MixMF and the results show that MixMF is the best performing method. TB406-2 PID: 4395 Adaptive Boosting: Dividing the Learning Set to Increase the Diversity and Performance of the Ensemble Joaqu??n Torres-Sospedra, Universitat Jaume I, Spain; Carlos Hern?andez-Espinosa, Universitat Jaume I, Spain; Mercedes Fern?andez-Redondo, Universitat Jaume I, Spain As shown in the bibliography, Boosting methods are widely used to build ensembles of neural networks. This kind of methods increases the performance with respect to a single network. Since Freund and Schapire introduced Adaptive Boosting in 1996 some authors have studied and improved Adaboost. In this paper we present Cross Validated Boosting a method based on Adaboost and Cross Validation. We have applied Cross Validation to the learning set in order to get an specific training set and validation set for each network. With this procedure the diversity increases because each network uses a specific validation set to finish its learning. Finally, we have performed a comparison among Adaboost and Crossboost on eight databases from UCI, the results show that Crossboost is the best performing method. TB406-3 PID: 4423 Ensemble of Competitive Associative Nets and A Method to Select Effective Number of Units Shuichi Kurogi, Kyushu Institute of Technology, JAPAN;
TB406
134
Shinya Tanaka, Kyushu Institute of Technology, JAPAN; Ryohei Koyama, Kyushu Institute of Technology, JAPAN; Takeshi Nishida, Kyushu Institute of Technology, JAPAN The competitive associative net called CAN2 has been shown effective in many applications, such as function approximation, control, rainfall estimation, time-series prediction, and so on, but the learning method has been constructed basically for reducing the training (empirical) error. In order to reduce prediction (generalization) error, we, in this article, try to apply the ensemble scheme to the CAN2 and present a method to select an effective number of units for the ensemble. We show the result of numerical experiments and examine the effectiveness of the present method. TB406-4 PID: 4917 Using Weighted Combination-based Methods in Ensembles with Different Levels of Diversity Thiago Dutra, Federal University of Rio Grande do Norte(UFRN), Brazil; Anne M P Canuto, Federal University of Rio Grande do Norte(UFRN), Brazil; Marcilo C P de Souto, Federal University of Rio Grande do Norte(UFRN), Brazil There are two main approaches to combine the output of classifiers within a multi-classifier system, which are: combination-based and selection-based methods. This paper presents an investigation of how the use of weights in some non-trainable simple combination-based methods applied to ensembles with different levels of diversity. It is aimed to analyse whether the use of weights can decrease the dependency of ensembles on the diversity of their members. TB406-5 PID: 4225 Integrated Neural Network Ensemble Algorithm Based on Clustering Technology Bingjie Liu, Xi??£¤an Institute of Hi-Tech, China; Changhua Hu, Xi??£¤an Institute of Hi-Tech, China Neural network ensemble (NNE) focuses on two aspects: how to generate component NNs and how to ensemble. The two interplayed aspects impact greatly on performance of NNE. Unfortunately, the two aspects were investigated separately in almost previous works. An integrated neural network ensemble (InNNE) is proposed in the paper, which was an integrated ensemble algorithm not only for dynamically adjusting weights of an ensemble, but also for generating component NNs based on clustering technology. InNNE classifies the training set into different subsets with clustering technology, which are used to train different component NNs. The weights of an ensemble are adjusted by the correlation of input data and the center of different training subsets. InNNE can increase the diversity of component NNs and decreases generalization error of ensemble. The paper provided both analytical and experimental evidence that support the novel algorithm.
TB407
135
TB407 Fuzzy Systems II Section Chair: Yi ZHOU, Nanyang Technological University TB407-1 PID: 5227 A Fuzzy LMS Neural Network Method for Evaluation of Importance of Indices in MADM Feng Kong, North China Electric Power University, China; Hongyan Liu, North China Electric Power University, China A fuzzy LMS (least-mean-square algorithm) neural network evaluation model, with fuzzy triangular numbers as inputs, is set up to compare the importance of different indices. The model can determine attribute or index weights (importance) automatically so that they are more objectively and accurately distributed. The model also has a strong self-learning ability so that calculations are greatly reduced and simplified. Further, decision maker’s specific preferences for uncertainty, i.e., risk-averse, risk-loving or risk-neutral, are considered in the evaluation model. Hence, our method can give objective results while taking into decision maker’s subjective intensions. Meanwhile, it is simple. A numerical example is given to illustrate the method. TB407-2 PID: 5232 Fuzzy RBF Neural Network Model for Multiple Attribute Decision Making Feng Kong, North China Electric Power University, China; Hongyan Liu, North China Electric Power University, China This paper studies how to compare and select one best alternative, from the new alternatives, according to historical or current ones. Previous methods not only need a lot of data but also are complex. So, we put forward an RBF neural network method that not only has the advantages of common neural network methods, but also needs much less samples and are straightforward. The number of neurons at the hidden level is easily determined. This model can determine attribute weights automatically so that weights are more objectively and accurately distributed. Further, decision maker’s specific preferences for uncertainty, i.e., risk-averse, risk-loving or risk-neutral, are considered in the determination of weights. Hence, our method can give objective results while taking into decision maker’s subjective intensions. A numerical example is given to illustrate the method. TB407-3 PID: 5257 A Brain-Inspired Fuzzy Semantic Memory Model for Learning and Reasoning with Uncertainty W.L. Tung, Nanyang Technological University, Singapore; C. Quek, Nanyang Technological University, Singapore Decision making pervades human experience. The human decision process is driven by rational reasoning, which is the capacity to use the faculty of reason to facilitate logical thinking and uncertain but sensible arguments from knowledge. Experience refers to the accumulation and the continuous neurological organization of knowledge via the repeated exposure to information and its effective usage. Conventional knowledge engineering
TB407
136
and decision support systems failed when they are applied to domains with uncertain information, where the uncertainties generally manifest as measurement noises. This paper presents T2-GenSoFNN, a brain-inspired fuzzy semantic memory model embedded with Type-2 fuzzy logic inference for learning and reasoning with noise-corrupted data. The proposed T2-GenSoFNN model is applied to the modeling of human insulin levels for the proper regulation of blood glucose concentration in diabetes therapy. The results are encouraging. TB407-4 PID: 4937 A Jumping Genes Paradigm with Fuzzy Rules for Optimizing Digital IIR Filters Sai-Ho Yeung, City University of Hong Kong, Hong Kong Kim-Fung Man, City University of Hong Kong, Hong Kong A Jumping Genes Paradigm that combines with fuzzy rules is applied for optimizing the digital IIR filters. The criteria that govern the quality of the optimization procedure are based on two basic measures. A newly formulated performance metric for the digital IIR filter is formed for checking its performance while its system order which usually reflects upon the required computational power is also adopted as another objective function for the optimization. The proposed scheme in this paper was able to obtain frequency-selective filters for lowpass, highpass, bandpass and bandstop with better performance than those previously obtained and the filter system order was also optimized with lower possible number.
TB408
137
TB408 Theoretical Model and Analysis II Section Chair: Manwai MAK, Hong Kong Polytechnic University TB408-1 PID: 5580 A solution to the Curse of Dimensionality Problem in Pairwise Scoring Techniques Man-Wai Mak, The Hong Kong Polytechnic University, Hong Kong; Sun-Yuan Kung, Princeton University, United States of America This paper provides a solution to the curse of dimensionality problem in the pairwise scoring techniques that are commonly used in bioinformatics and biometrics applications. It has been recently discovered that stacking the pairwise comparison scores between an unknown patterns and a set of known patterns can result in feature vectors with nice discriminative properties for classification. However, such technique can lead to curse of dimensionality because the vectors size is equal to the training set size. To overcome this problem, this paper shows that the pairwise score matrices possess a symmetric and diagonally dominant property that allows us to select the most relevant features independently by an FDA-like technique. Then, the paper demonstrates the capability of the technique via a protein sequence classification problem. It was found that 10-fold reduction in the number of feature dimensions and recognition time can be achieved with just 4% reduction in recognition accuracy. TB408-2 PID: 5302 First passage problem for the Ornstein-Uhlenbeck neuronal model C.F. Lo, The Chinese University of Hong Kong, Hong Kong; T.K. Chung, The Chinese University of Hong Kong, Hong Kong In this paper we propose a simple and efficient method for computing accurate estimates (in closed form) of the first passage time density of the Ornstein-Uhlenbeck neuronal model through a fixed boundary (i.e. the interspike statistics of the stochastic leaky integrate-and-fire neuron model). This new approach can also provide very tight upper and lower bounds (in closed form) for the exact first passage time density in a systematic manner. Unlike previous approximate analytical attempts, this novel approximation scheme not only goes beyond the linear response and weak noise limit, but it can also be systematically improved to yield the exact results. Furthermore, it is straightforward to extend our approach to study the more general case of a deterministically modulated boundary. TB408-3 PID: 5524 Delay-dependent Exponential Estimates of Stochastic Neural Networks with Time Delay Zhan Shu, The University of Hong Kong, Hong Kong; James Lam, The University of Hong Kong, Hong Kong This paper is concerned with the exponential estimating problem for a class of stochastic neural networks with time delay. A sufficient condition, which does not only guarantee the exponential stability but also gives the estimates of decay rate and decay coefficient, is established in terms of a new Lyapunov-Krasovskii functional and the linear matrix
TB408
138
inequality (LMI) technique. The estimating procedure is implemented by solving a set of LMIs, which can be checked easily by effective algorithms. A numerical example is provided to illustrate the effectiveness of the theoretical results. TB408-4 PID: 5467 Biased Minimax Probability Machine for Relevance Feedback in Image Retrieval Xiang Peng, The Chinese University of Hong Kong, Hong Kong In recent years, Minimax Probability Machine (MPM) has demonstrated excellent performance in a variety of pattern recognition problems. At the same time various machine learning methods have been used on relevance feedback tasks in Content-based Image Retrieval (CBIR). One of the problems in typical techniques for relevance feedback is that they treat the relevant feedback and irrelevant feedback equally. In other words, the negative instances largely outnumber the positive instances. Hence, the assumption that they are balanced is incorrect. In this paper we study how MPM can be applied to image retrieval, more precisely, Biased MPM during the relevance feedback iterations. We formulate the relevance feedback based on a modified MPM called Biased Minimax Probability Machine (BMPM). Different from previous methods, this model directly controls the accuracy of classification of the future data to build up biased classifiers. Hence, it provides a rigorous treatment on imbalanced data. Mathematical formulation and explanations are provided for showing the advantages. Experiments are conducted to evaluate the performance of our proposed framework, in which encouraging and promising experimental results are obtained. TB408-5 PID: 5051 Improvement of the Perfect Recall Rate of Block Splitting Type Morphological Associative Memory Using a Majority Logic Approach Takashi Saeki, Kyushu Institute of Technology, Japan; Tsutomu Miki, Kyushu Institute of Technology, Japan In this paper, a new improvement approach of the perfect recall rate of a block splitting type morphological associative memory (BMAM) is presented. The BMAM is one of MAMs without the kernel image, which is realized in more compact size as keeping the perfect recall rate as same as a normal MAM (without the kernel image). However, the MAM without kernel image has a problem that the perfect recall rate is inferior to a standard MAM (with the kernel image). Therefore, we try to improve the problem by a majority logic scheme and confirm the effectiveness of the proposed approach through autoassociation experiments of alphabet patterns compared to the traditional approaches in terms of the noise tolerance.
TC401
139
Thursday Session C, 5 October 2006
TC401 Invited talk by Aike Guo Invited talk by Tianyou Chai Section Chair: Jun WANG, The Chinese University of Hong Kong
TC402
140
TC402 Cognitive Processing / Learning IV Section Chair: Qi ZHANG, Sony Computer Science Laboratories TC402-1 PID: 4735 Semantic Addressable Encoding Cheng-Yuan Liou, National Taiwan University; Jau-Chi Huang, National Taiwan University; Wen-Chie Yang, National Taiwan University; This paper presents an automatic acquisition process to acquire the semantic meaning for the words. This process obtains the representation vectors for stemmed words by iteratively improving the vectors, using a trained Elman network. Experiments performed on a corpus composed of Shakespeare’s writings show its linguistic analysis and categorization abilities. TC402-2 PID: 4971 A Time-Dependent Model of Information Capacity of Visual Attention Xiaodi Hou, Shanghai Jiao Tong University, China; Liqing Zhang, Shanghai Jiao Tong University, China; What does a human’s eye tell a human’s brain? In this paper, we analyze the information capacity of visual attention. Our hypothesis is that the limit of perceptible spatial frequency is related to observing time. Given more time, one can obtain higher resolution - that is, higher spatial frequency information, of the presented visual stimuli. We de- signed an experiment to simulate natural viewing conditions, in which time dependent characteristics of the attention can be evoked; and we recorded the temporal responses of 6 subjects. Based on the experiment results, we propose a person-independent model that characterizes the behavior of eyes, relating visual spatial resolution with the duration of attentional concentration time. This model suggests that the information capacity of visual attention is time-dependent. TC402-3 PID: 4657 A neural model for stereo transparency with the population of hybrid energy models Osamu Watanabe, Muroran Institute of Technology, Japan; The disparity energy model can explain physiological properties of binocular neurons in early visual cortex quantitatively. Therefore, many physiologically-plausible models for binocular stereopsis employed the disparity energy model as a model neuron. These models can explain a variety of psychological data. However, most of them cannot handle with stereo transparency. Here, we develop a simple model for transparency perception with the disparity energy model, and examine the ability to detect overlapping disparities. Computer simulations showed that the model properties of transparency detection are consistent with many psychophysical findings. TC402-4 PID: 4422
TC402
141
Application of Competitive Associative Nets to Plane Extraction from Range Data Shuichi Kurogi, Kyushu Institute of Technology, Japan; Shota Okada, Kyushu Institute of Technology, Japan; Shingo Inoue, Kyushu Institute of Technology, Japan; Takeshi Nishida, Kyushu Institute of Technology, Japan; This article describes an application of competitive associative net called CAN2 to plane extraction from 3D range images measured by a laser range filter (LRF). The CAN2 basically is a neural net which learns efficient piecewise linear approximation of nonlinear functions, and in this application it is utilized for learning piecewise planner (linear) surfaces from the range data. As a result of the learning, the obtained piecewise planner surfaces are more precise than the actual planner surfaces, so that we introduce a method to gather piecewise planner surfaces for reconstructing the actual planner surfaces. We apply the method to the real range data, and examine the effectiveness and the performance. TC402-5 PID: 5367 Representation of 3-D Volumetric Object from the Pantomime Effect and Shading Cues in Human Brain Qi Zhang, Sony Computer Science Laboratories, Inc., Japan; Ken Mogi, Sony Computer Science Laboratories, Inc., Japan; Human ability to process visual information of outside world is yet far away to approach for man-made systems in the high accuracy and speed. In particular, human beings can perceive 3-D object from various cues, such as binocular disparity and monocular shading cues. Understanding of the mechanism of human visual processing will lead to a breakthrough in creating artificial visual systems. Here, we study the human 3-D volumetric object perception that is induced by a visual phenomenon named as the pantomime effect and by the monocular shading cues. We measured human brain activities using fMRI when the subjects were observing the visual stimuli. A coordinated system of brain areas, including those involved in semantic processing, working memory, spatial processing, in addition to the occipital visual areas was found to be involved in the volumetric object perception.
TC403
142
TC403 Hardware Implementation Section Chair: Toshi SAITO, Hosei University TC403-1 PID: 5435 FPGA Discrete Wavelet Transform Neural Network Activate Functions and Encoder/Decoder Implementation Pedro Henrique Cox, Fundac?o Universidade Federal de Mato Grosso do Sul, Brasil Aparecido Augusto de Carvalho, Universidade do Estado de S?o Paulo, Brasil In a multi-input and multi-output feedforward wavelet neural network, orthogonal wavelet basis functions are used as activate function instead of sigmoid function of feedforward network. This paper addresses the solution on processing biological data such as cardiac beats, audio and ultrasonic range, calculating wavelet coefficients in real time, with processor clock running at frequency of present ASIC’s and FPGA. The Parallel Filter Architecture for DWT has been improved, calculating wavelet coefficients in real time with hardware reduced up to 60%. The new architecture, which also processes IDWT, is implemented with the Radix-2 or the Booth-Wallace Constant multipliers. One integrated circuit Encoder/Decoder, ultrasonic range, is presented. TC403-2 PID: 5218 Synchronization via multiplex spike-trains in digital pulse-coupled networks Takahiro KABE, Hosei University, Japan Hiroyuki TORIKAI, Hosei University, Japan Toshimichi SAITO, Hosei University, Japan This paper studies pulse-coupled network of digital spiking neurons and its basic dynamics. The neuron is constructed by coupling two shift registers and has a variety of spike-trains which correspond to digital codes through an inter-spike interval (ISI) modulation. The pulsecoupled network has master-slave configuration. All the spike-trains of neurons in the master side are multiplexed additionally and are transmitted to the slave side via single line. Neurons in the slave side are connected by dynamic winner-take-all function. As parameters are selected suitably, the slave can realize demultiplexing and master-slave synchronization is achieved. VHDL simulation is also discussed for FPGA implementation and this digital network is compared with an analog network. TC403-3 PID: 4956 Application of Artificial Neural Network to Modeling and Simulation of On-Chip Interconnects Wei Zhou, Wuhan University, China In modern very large-scale integration (VLSI) technology, the operating frequencies are fast reaching the vicinity of gigahertz (GHz) and switching times are getting to the sub-nanosecond or nanosecond levels. The ever increasing quest for high-speed applications is placing higher demands on interconnect performances, and has highlighted the previously negligible effects of interconnects. Traditional methods used to model and
TC403
143
simulate high-speed interconnects are highly CPU-intensive thus are not feasible for on-line use in large-scale CAD and optimizing techniques. In this paper, a feed forward neural network based approach is presented, which acts as a black box between the physical/geometry information and electrical parameters of a certain interconnect network. Numerical experiments results show that the proposed approach is effective both in computation time and accuracy.
TC404
144
TC404 Pattern Classification/Classifiers & Classification IV (Face Analysis & Processing) Section Chair: Christopher YANG, The Chinese University of Hong Kong TC404-1 PID: 5013 A Novel Model for Gabor-based Independent Radial Basis Function Neural Networks and Its Application to Face Recognition GaoYun An, Beijing Jiaotong University, China QiuQi Ruan, Beijing Jiaotong University, China In this paper, a novel model for Gabor-based independent radial basis function (IRBF) neural network is proposed and applied to face recognition. In the new model, a bank of Gabor filters is first built to extract Gabor face representations characterized by selected frequency, locality and orientation to cope with various illuminations, facial expression and poses in face recognition. Then principal component analysis (PCA) is adopted to reduce the dimension of the extracted Gabor face representations for every face sample. At last, a new IRBF neural network is built to extract high-order statistical features of extracted Gabor face representations with lower dimension and to classify these extracted high-order statistical features. According to the experiments on the famous CAS-PEAL face database, our proposed approach could outperform ICA with architecture II (ICA2) and kernel PCA (KPCA) with standing testing sets proposed in the current release disk of the CAS-PEAL face database. TC404-2 PID: 5421 Fast Competition Approach using Self Organizing Map for Multivariate Data Applications Alaa Sagheer, Kyushu University, Japan Nayouki Tsuruta, Fukuoka University, Japan Rin-Ichiro Taniguchi, Kyushu University, Japan Sakashi Maeda, Fukuoka University, Japan Though, the Self Organizing Map is one of the most widely used neural network paradigm based on unsupervised competitive learning, its competition “search” algorithm is slow when the size of the map is large. In this paper, we present a new strategy capable to reduce SOM’s computation efforts. The new SOM can works effectively as a feature extractor for all kinds of manifolds of the limited or multivariate data sets; even in the curved manifolds. Three data sets are utilized to illustrate how the proposed algorithm reduces significantly the computation efforts (or time) of SOM effectively and maintains the same, or may be better, recognition accuracy of original SOM. For N-dimensions feature space, it is shown here that the computation effort to get the best matching units is reduced to O(D1+ D2+...+ DN) instead of O(D1 x D2 x ... x DN), where Di is the number of neurons through the dimension i. TC404-3 PID: 4877 An efficient Pseudoinverse Linear Discriminant Analysis method and its application to face recognition with great variations in expression, lighting and occlusion
TC404
145
Jun Liu, Nanjing University, China Songcan Chen, Nanjing University, China Daoqiang Zhang, Nanjing University, China Xiaoyang Tan, Nanjing University, China Pseudoinverse Linear Discriminant Analysis (PLDA) is a classical and pioneer method that deals with the Small Size Sample (SSS) problem in LDA when applied to such application as face recognition. However, it is expensive in computation and storage due to manipulating on extremely large dXd matrices, where d is the dimensionality of the sample image. As a result, although frequently cited in literature, PLDA is hardly compared in terms of classification performance with the newly proposed methods. In this paper, we propose a new feature extraction method named RSw+LDA, which is 1) much more efficient than PLDA in both computation and storage; and 2) theoretically equivalent to PLDA, meaning that it produces the same projection matrix as PLDA. Our experimental results on AR face dataset, a challenging dataset with variations in expression, lighting and occlusion, show that PLDA (or RSw+LDA) can achieve significantly higher classification accuracy than the recently proposed Linear Discriminant Analysis via QR decomposition and Discriminant Common Vectors. TC404-4 PID: 5440 Multiple Decision Templates with Adaptive Features for Fingerprint Classification Junki Min, Yonsei University, Korea Sung-Bae Cho, Yonsei University, Korea This paper proposes a novel fingerprint classification algorithm using multiple decision templates of support vector machines (SVMs) with adaptive features. In order to overcome the intra class ambiguity of fingerprints, the proposed algorithm extracts the feature vectors from adaptively detected feature region and classifies these adaptive features using multiple decision templates which are the clustered outputs of SVMs. Experimental results on NIST4 fingerprint database reveal the effectiveness and validity of our algorithm for fingerprint classification.
TC405
146
TC405 PSO II-Applications Section Chair: Sanghoun OH, Gwangju Institute of Science and Technology TC405-1 PID: 4282 Extended Particle Swarm Optimiser with Adaptive Acceleration Coefficients and Its Application in Nonlinear Blind Source Separation Ying Gao, Guangzhou University, China; Zhaohui LI, Guangzhou University, China; Hui Zheng, Guangzhou University, China; Huailiang Liu, Guangzhou University, China; First, based on the particle swarm optimization, an extended particle swarm optimizer with acceleration coefficients (EPSO_AAC) is presented. The personal best particle is replaced by the average of personal best particles in swarm at generation, and time-varying acceleration coefficients are applied by establishing a nonlinear functional relationship between acceleration coefficients and the different of the average fitness of all particles and the fitness of the global best particle. The proposed algorithm uses more particles’ information, and adjusts adaptively “cognition” component and “social” component by time-varying acceleration coefficients, thus improves convergence performance. Then, the proposed algorithm is applied to nonlinear blind source separation. The demixing system of the nonlinear mixtures is modeled using a multi-input multi-output B-spline neural network whose weights are optimized under the criterion of independence of its outputs by EPSO_AAC. The experiment results demonstrate that the proposed algorithms are effective, and have good convergence performance. TC405-2 PID: 5288 PSO-Based Hyper-parameters Selection for LS-SVM Classifiers X. C. Guo, Jilin University, Northeast Dianli University, China; Y. C. Liang, Jilin University, China; C. G. Wu, Jilin University, Beijing Jiaotong University, China; C. Y. Wang, Jilin University, China; The determination for hyper-parameters including kernel parameters and the regularization is important to the performance of least squares support vector machines (LS-SVMs). In this paper, the problem of model selection for LS-SVMs is discussed. The particle swarm optimization (PSO) is introduced to select the LS-SVMs hyper-parameters. In the proposed method we do not need to consider the analytic property of the generalization performance measure and the number of hyper-parameters. The feasibility of this method is evaluated on benchmark data sets. Experimental results show that better performance can be obtained. Moreover, different kinds of kernel families are investigated by using the proposed method. Experimental results also show that the best and good test performance could be obtained by using the SRBF and RBF kernel functions, respectively. TC405-3 PID: 5075
TC405
147
Dimensionality Reduction for Evolving RBF with Particle Swarms Junying Chen, Xian JiaoTong University,China; Zheng Qin, Xian JiaoTong University,China; Dimensionality reduction including both feature selection and feature extraction techniques are useful for improving the classification accuracy of neural network. In this paper, particle swarm optimization (PSO) algorithm was proposed for simultaneous feature extraction and feature selection. First PSO was used to simultaneous feature extraction and selection in conjunction with k-nearest-neighbor (k-NN) for individual fitness evaluation. With the derived feature set based on the first step, PSO was then used to evolve RBF networks dynamically. Experimental results on four datasets show that RBF networks evolved by our proposed algorithm have more simple architecture and stronger generalization ability with the same as, or even better classification performance when compared with the networks evolved by other methods.
TC406
148
TC406 Forecasting & Prediction II Section Chair: Xun LIANG, Peking University TC406-1 PID: 4472 A Distributed Computing Service for Neural Networks and its Application to Flood Peak Forecasting Jun Zhu, Chinese Academy of Sciences, China Chunbo Liu, Peking University, China Jianhua Gong, Chinese Academy of Sciences, China Daojun Wang, Chinese Academy of Sciences, China Tao Song, Inha University, Korea How to exploit current information techniques for rapidly and accurately building a fittest neural network becomes increasingly significant for flood peak forecasting. This paper firstly designs a distributed computing architecture and builds a computing environment based on Grid technologies. Then a distributed computing service for neural networks based on a genetic algorithm and a modified BP algorithm is designed and developed to rapidly and accurately building a fittest neural network for flood peak forecasting. Finally, a distributed computing prototype system is developed and implemented on a case study of the flood prevention in Shenzhen city, China. The experiment result shows that the scheme addressed in the paper is efficient and feasible. TC406-2 PID: 5137 Automatic inference of cabinet approval ratings by information-theoretic learning Ryotaro Kamimura, Information Science Laboratory, Japan Fumihiko Yoshida, Tokai University, Japan In this paper, we demonstrate that cabinet approval ratings can automatically be inferred with good performance by a neural network technique, that is, information-theoretic competitive learning. Because cabinet approval rating estimation is an extremely complex process with much non-linearity, neural networks may give much better performance than conventional statistical methods. Though an attempt to infer public opinions seems to be a challenging topic for machine learning, little attempts have been made to infer approval ratings to our best knowledge. In this context, we try to apply information-theoretic competitive learning to the problem of cabinet approval ratings. Information-theoretic competitive learning has been developed so as to simulate competitive processes of neurons. One of the main characteristics of the method is that it is a very soft-type of competitive learning in which conventional competitive learning is only a special case. Though the method seems to be promising due to its general property, we have had a few experimental results to show better performance. Experimental results show that without any teacher information neural networks can appropriately infer the rise and fall of approval ratings through a process of information maximization. This experiment result surely opens up new perspectives for neural networks as well as mass communication studies. TC406-3 PID: 5047
TC406
149
Short-Term Load Forecasting using Multiscale BiLinear Recurrent Neural Network with an Adaptive Learning Algorithm Dong-Chul Park, Myong Ji University, Korea Chung Nguyen Tran, Myong Ji University, Korea Yunsik Lee, Korea Electronics Tech. Inst., Seongnam, Korea In this paper, a short-term load forecasting model using a Multiscale BiLinear Recurrent Neural Network with an adaptive learning algorithm (M-BLRNN(AL)) is proposed. The proposed M-BLRNN(AL) model is based on a wavelet-based neural network architecture formulated by a combination of several individual BLRNN models. The wavelet transform adopted in the M-BLRNN(AL) is employed to decompose the load curve into a mutiresolution representation. Each individual BLRNN model is used to forecast the load signal at each resolution level obtained by the wavelet transform. The learning process is further improved by applying an adaptive learning algorithm at each resolution level. Experiments and results on load data from the North-American Electric Utility (NAEU) show that the proposed M-BLRNN(AL) model converges faster and archives better forecasting performance in comparison with other conventional models. TC406-4 PID: 4284 A New Approach to Load Forecasting: Using Semi-Parametric Method and Neural Networks Abhisek Ukil, Tshwane University of Technology, South Africa Jaco Jordaan, Tshwane University of Technology, South Africa A new approach to electrical load forecasting is investigated. The method is based on the semi-parametric spectral estimation method that is used to decompose a signal into a harmonic linear signal model and a non-linear part. A neural network is then used to predict the nonlinear part. The final predicted signal is then found by adding the neural network predicted non-linear part and the linear part. The performance of the proposed method seems to be more robust than using only the raw load data.
TC407
150
TC407 Data Preprocessing/Dimension Reduction II Section Chair: Zenglin XU, The Chinese University of Hong Kong TC407-1 PID: 5264 An Excellent Feature Selection Model using Gradient-based and Point injection techniques D. Huang, City University of Hong Kong, Hong Kong Tommy W. S. Chow, City University of Hong Kong, Hong Kong This paper focuses on enhancing the effectiveness of filter feature selection models from two aspects. One is to modify feature searching engines based on optimization theory, and the other is to improve the regularization capability using point injection techniques. The second topic is undoubtedly important in the situations where overfitting is likely to be met, for example, the ones with only small sample sets available. Synthetic and real data are used to demonstrate the contribution of our proposed strategies. TC407-2 PID: 5295 Driven Forward Features Selection: a comparative study on Neural Networks Vincent Lemaire,France Telecom R&D Lannion, France Raphael F?eraud, France Telecom R&D Lannion, France In the field of neural networks, feature selection has been studied for the last ten years and classical as well as original methods have been employed. This paper reviews the efficiency of four approaches to do a driven forward features selection on neural networks. We assess the efficiency of these methods compare to the simple Pearson criterion in case of a regression problem. TC407-3 PID: 5376 Non-negative Matrix Factorization based Text Mining: Feature extraction and Classification P C Barman, Korea Advanced Institute of Science and Technology, Korea Nadeem Iqbal, Korea Advanced Institute of Science and Technology, Korea Soo-Young Lee, Korea Advanced Institute of Science and Technology, Korea The unlabeled document or text collections are becoming larger and larger which is common and obvious; mining such data sets are a challenging task. Using the simple word-document frequency matrix as feature space the mining process is becoming more complex. The text documents are often represented as high dimensional about few thousand sparse vectors with sparsity about 95 to 99% which significantly affects the efficiency and the results of the mining process. In this paper, we propose the two-stage Non-negative Matrix Factorization (NMF): in the first stage we tried to extract the uncorrelated basis probabilistic document feature vectors by significantly reducing the dimension of the feature vectors of the word-document frequency from few thousand to few hundred, and in the second stage for clustering or classification. In our proposed approach it has been observed that the clustering or classification performance with more
TC407
151
than 98.5% accuracy. The dimension reduction and classification performance has observed for the Classic3 dataset. TC407-4 PID: 4986 Adaptive Parameters Determination Method of Pulse Coupled Neural Network Based on Water Valley Area Min Li, Xi’an Jiaotong University, China Wei Cai, Xi’an Research Inst. of Hi-Tech, China Zheng Tan, Xi’an Jiaotong University, China Pulse coupled neural network (PCNN) is different from traditional artificial neural networks, models of which have biological background and are based on the experimental observations of synchronous pulse bursts in the cat visual cortex. However, it is very difficult to determine the exact relationship between the parameters of PCNN model. Focusing on the famous difficult problem of PCNN, how to determine the optimum parameters automatically, and this paper proposes the definition of water valley area, establishes a modified PCNN, and puts forward an adaptive PCNN parameters determination algorithm based on water valley area. Extensive experimental results on image processing demonstrate its validity and robustness. TC407-5 PID: 5087 Feature Selection Based on Minimum Error Minimax Probability Machine Zenglin Xu, The Chinese University of Hong Kong, Hong Kong Michael R. Lyu, The Chinese University of Hong Kong, Hong Kong Feature selection is an important task in pattern recognition. Support Vector Machine (SVM) and Minimax Probability Machine (MPM) have been successfully used as classification framework for feature selection. However, these paradigms cannot automatically control the balance between prediction accuracy and the number of selected features. The selected feature subsets are also not stable in different partitions and different runs. Minimum Error Minimax Probability Machine (MEMPM) has been proposed for classification recently. In this paper, we outline MEMPM to select optimal feature subset with good stability and automatic balance between prediction accuracy and the size of feature subset. The experiments against feature selection with SVM and MPM show the advantages of MEMPM formulation in stability and automatic balance between feature subset size and prediction accuracy.
TC408
152
TC408 Theoretical Modeling and Analysis III Section Chair: Tingwen HUANG, Texas A&M University at Qatar TC408-1 PID: 4969 Two Methods for Sparsifying Probabilistic Canonical Correlation Analysis Colin Fyfe, The University of Paisley, Scotland Gayle Leen, The University of Paisley, Scotland We have recently developed several ways of performing Canonical Correlation Analysis [1, 5, 7, 4] with probabilistic methods rather than the standard statistical tools. However, the computational demand of training such methods scales with the square of the number of samples, making these methods uncompetitive with e.g. artificial neural network methods. In this paper, we examine two recent developments which sparsify probabilistic methods of performing canonical correlation analysis. TC408-2 PID: 4886 Turbo decoding as an instance of EM algorithm Saejoon Kim, Sogang University, Korea The Baum-Welch algorithm is a technique for the maximum likelihood parameter estimation of probabilistic functions of Markov processes. We apply this technique to nonstationary Markov processes and explore a relationship between the Baum-Welch algorithm and the BCJR algorithm. Furthermore, we apply the Baum-Welch algorithm to two nonstationary Markov processes and obtain the turbo decoding algorithm. TC408-3 PID: 4817 Dynamical behaviors of a large class of delayed differential systems with discontinuous right-hand side Wenlian Lu, Fudan University, China Tianping Chen, Fudan University, China In this paper, we investigate dynamical behaviors of a large class of delayed differential systems with discontinuous right-hand side. This class of delayed differential systems includes Hopfield and Cellular neural networks with discontinuous activations as special cases. We prove that under some mild conditions, this system has a unique almost- periodic solution, which is globally exponential stable. TC408-4 PID: 5183 Reinforcement Learning Algorithm with CTRNN in Continuous Action Space Hiroaki Arie, Waseda University, Japan Jun Namikawa, RIKEN, Brain Science Institute, Japan Tetsuya Ogata, Kyoto University, Japan Jun Tani, RIKEN, Brain Science Institute, Japan Shigeki Sugano, Waseda University, Japan
TC408
153
There are some difficulties in applying traditional reinforcement learning algorithms to motion control tasks of robot. It is because most algorithms are concerned with discrete actions and based on the assumption of complete observability of the state. This paper deals with these two problems by combining the reinforcement learning algorithm and CTRNN learning algorithm. We carried out an experiment on the pendulum swing-up task without rotational speed information. It is shown that the information about the rotational speed, which is considered as a hidden state, is estimated and encoded on the activation of a context neuron. As a result, this task is accomplished in several hundred trials using the proposed algorithm. TC408-5 PID: 4720 Entropy based Associative Model Masahiro Nakagawa, Nagaoka University of Technology In this paper, an entropy based associative memory model will be proposed and applied to memory retrievals with an orthogonal learning model to compare with the conventional autoassociative model with a quadratic Lyapunov functions. In the present approach, the updating dynamics will be constructed on the basis of the entropy minimization strategy which may be reduced asymptotically to the autocorrelation dynamics as a special case. From numerical results, it will be found that the presently proposed novel approach realizes the larger memory capacity in comparison with the autocorrelation model based on dynamics such as association according to the higher-order correlation involved in the proposed dynamics.
TP
154
Thursday Poster Session II, 5 October 2006
TP-1 PID: 4159 Ultrasound Image Segmentation by Using Wavelet Transform and Self-organizing Neural Network Zafer scan, Istanbul Technical University, Turkey Mehmet Nadir Kurnaz, Istanbul Technical University, Turkey Z¨¹mray Dokur, Istanbul Technical University, Turkey Tamer ?lmez, Istanbul Technical University, Turkey This paper presents an improved incremental self-organizing map (I2SOM) network that uses automatic threshold (AT) value for the segmentation of ultrasound (US) images. In order to show the validity of proposed scheme, it has been compared with Kohonen’s SOM. Two dimensional (2D) fast Fourier transform (FFT) and 2D continuous wavelet transform (CWT) were computed in order to form the feature vectors of US bladder and phantom images. In this study, it is observed that the proposed automatic threshold scheme for ISOM network has significantly eliminated the former ISOM network’s threshold problem for US images. This improvement enhances the robustness of ISOM algorithm. Obtained results show that ISOM with AT value has similar segmentation performance with Kohonen’s network. TP-2 PID: 4162 Genetically Designed Time Delay Neural Networks for Multiple-interval Urban Freeway Traffic Flow Forecasting Ming Zhong, University of New Brunswick, Canada Satish Sharma, University of Regina Pawan Lingras, Saint Mary¡¯s University, Canada Previous research for short-term traffic prediction mostly forecasts only one time interval ahead. Such a methodology may not be adequate for response to emergency circumstances and road maintenance activities that last for a few hours or a longer period. In this study, various approaches, including naïve factor methods, exponential weighted moving average (EWMA), autoregressive integrated moving average (ARIMA), and genetically designed time delay neural network (GA-TDNN) are proposed for predicting traffic flow of continuous 12 hours ahead on a freeway near the City of Calgary, Canada. Study results show that the ARIMA models outperform EWMA models, which in turn superior to the factor methods. GA-TDNN results in only comparable accuracy with the ARIMA model, and it seems not worth to develop such complicated models. However, the adaptive nature of neural networks promises better accuracy as they are exposed to more observations during field operation. Its non-parametric approach also guarantees a greater portability and much faster computing speed for real-time applications. TP-3 PID: 4239 A New Subspace Analysis Approach Based on Laplacianfaces Yan Wu, Tongji University, China Ren-Min Gu, Tongji University, China
TP
155
A new subspace analysis approach named ANLBM is pro- posed based on Laplacianfaces. It uses the discriminant information of training samples by supervised mechanism, enhances within-class local information by an objective function. The objective function is used to construct adjacency graph’s weight matrix. In order to avoid the drawback of Laplacianfaces’ PCA step, ANLBM uses kernel mapping. ANLBM changes the problem from minimum eigenvalue solution to maximum eigenvalue solution, reduces the redundancy of the computing and increases the precision of the result. The experiments are performed on ORL and Yale databases. Experimental results show that ANLBM has a better performance. TP-4 PID: 4245 Implicit elitism in genetic search A. K. Bhatia, Banaras Hindu University, India S. K. Basu, Banaras Hindu University, India We introduce a notion of implicit elitism derived from the mutation operator in genetic algorithms. Probability of mutation less than 1/l ( l being the chromosome size) along with probability of crossover less than one induces implicit elitism in genetic search. It implicitly transfers a few chromosomes with above-average fitness unperturbed to the population at next generation, thus maintaining the progress of genetic search. Experiments conducted on one-max and 0/1 knapsack problems testify its efficacy. Implicit elitism in combination with traditional explicit elitism enhances the search capability of genetic algorithms. TP-5 PID: 4283 Practical Denoising of MEG Data using Wavelet Transform Abhisek Ukil, Tshwane University of Technology, South Africa Magnetoencephalography (MEG) is an important noninvasive, nonhazardous technology for functional brain mapping, measuring the magnetic fields due to the intracellular neuronal current flow in the brain. However, the inherent level of noise in the data collection process is large enough to obscure the signal(s) of interest most often. In this paper, a practical denoising technique based on the wavelet transform and the multiresolution signal decomposition technique is presented. The proposed technique is substantiated by the application results using three different mother wavelets on the recorded MEG signal. TP-6 PID: 4387 Genetic Algorithm for Satellite Customer Assignment S. S. Kim, Kangwon National University, Korea H. J. Kim, Kangwon National University, Korea V. Mani, Indian Institute of Science, India C. H. Kim, Kangwon National University, Korea The problem of assigning customers to satellite channels is considered. Finding an optimal allocation of customers to satellite channels is a difficult combinatorial optimization problem and is shown to be NP-complete in an earlier study. We propose a
TP
156
genetic algorithm (GA) approach to search for the best/optimal assignment of customers to satellite channels. Various issues related to genetic algorithms such as solution representation, selection methods, genetic operators and repair of invalid solutions are presented. A comparison of this approach with the standard optimization method is presented to show the advantages of this approach in terms of computation time. TP-7 PID: 4406 An Improved Primal-Dual Genetic Algorithm for Optimization in Dynamic Environments Hongfeng Wang, Northeastern University, China Dingwei Wang, Northeastern University, China Inspired by the complementary and dominance mechanism in nature, the Primal-Dual Genetic Algorithm (PDGA) has been proved successful in dynamic environments. In this paper, an important operator in PDGA, primal-dual mapping, is discussed and a new statistics-based primal-dual mapping scheme is proposed. The experimental results on the dynamic optimization problems generated from a set of stationary benchmark problems show that the improved PDGA has stronger adaptability and robustness than the original for dynamic optimization problems. TP-8 PID: 4418 Predication of Properties of Welding Joints Based on Uniform Designed Neural Network Shi Yu, Lanzhou Univ. of Tech., China Li Jianjun, Lanzhou Univ. of Tech., China Fan Ding, Lanzhou Univ. of Tech., China Chen Jianhong, Lanzhou Univ. of Tech., China It is difficult to predict the mechanical properties of welded joints because of non-linearity in welding process and complicated mutual effects in multi composition welding material. Based on these practical problems, the application of neural network technology in predicting mechanical properties of welding joints is developed. The modeling method has been studied and the author puts forward that the parameters of neutral network can be optimized by the method of uniform design. The neutral network model of mechanical properties of welding joints is established on the basis of the data of welding thermal simulation, and the experimental results show that this model can predict the mechanical properties including impact toughness, tensile strength, subdued strength, reduction ratio of area and hardness more accurately. At the same time, using this method can improve estimating precision largely compared with using traditional statistic method. That is, this method provides an effective approach to estimate the mechanical properties of welding joints. TP-9 PID: 4556 The Optimal Solution of TSP using the new Mixture initialization and sequential transformation method in Genetic Algorithm Rae-Goo Kang, The Chosun University of Korea, Korea Chai-Yeoung Jung, The Chosun University of Korea, Korea
TP
157
TSP(Traveling Salesman Problem) used widely for solving the optimization is the problem to find out the shortest distance out of possible courses where one starts a certain city, visits every city among N cities and turns back to a staring city. At this time, the condition is to visit N cities exactly only once. TSP is defined easily, but as the number of visiting cities increases, the calculation rate increases geometrically. This is why TSP is classified into NP-Hard Problem. Genetic Algorithm is used representatively to solve the TSP. Various operators have been developed and studied until now for solving the TSP more effectively. This paper applied the new Population Initialization Method (using the Random Initialization method and Induced Initialization method simultaneously), solved TSP more effectively, and proved the improvement of capability by comparing this new method with existing methods. TP-10 PID: 4655 Neural Virtual Sensor for the Inferential Prediction in Olefin Polymerization Xinggao Liu, Zhejiang University,China Youxian Sun, Zhejiang University,China Melt Index (MI) is considered an important quality variable determining product specification in olefin polymerization. In this paper, a novel neural soft-sensor architecture combined with Independent Component Analysis (ICA) and Multi-scale Analysis (MSA) is developed to infer the MI of polypropylene from other process variables. The simplified models without ICA or MSA are simultaneously developed as the basis of research. The detailed comparative researches are further carried out. The research results show that the proposed method provides promising prediction reliability and accuracy. TP-11 PID: 4794 Condition Surveillance for Plant Rotating Machinery Using Fuzzy Neural Network Tetsuro MITOMA, Mie University, Japan Peng CHEN, Mie University, Japan Huaqing WANG, Mie University, Japan Condition surveillance of rotating machinery in a plant is very important for guaranteeing the production efficiency and the plant safety. In a large scale plat, because there are enormous numbers of rotating machine, the condition surveillance for all rotating machines is not only time consuming and labor intensive, but also does not ensure the accuracy of condition judgments. These difficulties may cause serious machine accidents and bring great production losses. In order to improve the efficiency of the condition surveillance and detect faults at an early stage, this paper proposes a method of condition surveillance for plant rotating machinery using “Partially-linearized Neural Network (PLNN)” by which the state of a rotating machine can be automatically judged on the basis of the possibility of normal and abnormal states. When input rotating speed, power, shaft diameter and vibration parameter of a rotating machine into the PLNN, which has been finished learning, it will output the possibilities of normal and abnormal states for the inspected rotating machine. Practical example of condition surveillance for plant rotating machinery will verify that the method is effective. TP-12
TP
158
PID: 4836 Language Learning for the Autonomous Mental Development of Conversational Agents Jin-Hyuk Hong, Yonsei University, Korea Sungsoo Lim, Yonsei University, Korea Sung-Bae Cho, Yonsei University, Korea Since the manual construction of our knowledge-base has several crucial limitations when applied to intelligent systems, mental development has been investigated in recent years. Autonomous mental development is a new paradigm for developing autonomous machines, which are adaptive and flexible to the environment. Language development, a kind of mental development, is an important aspect of intelligent conversational agents. In this paper, we propose an intelligent conversational agent and its language development mechanism by putting together five promising techniques; Bayesian networks, pattern matching, finite state machines, templates, and genetic programming. Knowledge acquisition implemented by finite state machines and templates, and language learning by genetic programming are developed for language development. Several illustrations and usability tests show the usefulness of the proposed developmental conversational agent. TP-13 PID: 4842 Evolutionary approach to the game of checkers Magdalena Kusiak, Warsaw University of Technology, Poland Karol Wal?edzik, Warsaw University of Technology, Poland Jacek Ma?ndziuk, Warsaw University of Technology, Poland A new method of genetic evolution of linear and nonlinear heuristic evaluation functions in the game of checkers is presented and tested in the paper. Several practical issues concerning behavior of evolutionary methods in this task are pointed out and discussed. Experimental results confirm that proposed approach leads to relatively strong heuristics comparable to the ones used in some of commercial applications. TP-14 PID: 4875 Image Fusion Based on PCA and Undecimated Discrete Wavelet Transform Wei Liu, Information Science and Technology Institute, China Jie Huang, Information Science and Technology Institute, China Yongjun Zhao, Information Science and Technology Institute, China On the basis of analyzing the performances of popular image fusion methods, a new remote sensing image fusion method based on principal component analysis (PCA), high pass filter (HPF) and undecimated discrete wavelet transform (UDWT) is proposed. Some measure parameters are suggested to evaluate the fusion method. Experiments have been performed with the SPOT panchromatic image and the TM multi-spectral image. Both subjectively qualitative analysis and objectively quantitative evaluation verify the performance of the new method. With the same wavelet transform level, the fusion image using the proposed method preserves more sophisticated spatial details and distorts less spectral information in comparison with the fusion image using the traditional discrete wavelet transform (DWT) method.
TP
159
TP-15 PID: 4884 Stability of Periodic Solution in Fuzzy BAM Neural Networks with Finite Distributed Delays Tingwen Huang, Texas A&M University at Qatar, USA In this paper, we investigate fuzzy bidirectional associative memory (BAM) neural networks with finite distributed delays. Easily verifiable sufficient conditions for global exponential periodicity of fuzzy BAM neural networks with finite distributed delays are obtained. TP-16 PID: 4959 Retrieval-aware image compression, its format and viewer based upon learned bases Naoto Katsumata, Yahoo Japan, Japan Yasuo Matsuyama, Waseda University, Japan Takeshi Chikagawa, Nomura Research Institute, Japan Fuminori Ohashi, Waseda University, Japan Fumiaki Horiike, Waseda University, Japan Shun¡¯ichi Honma, Waseda University, Japan Tomohiro Nakamura, Waseda University, Japan A retrieval-aware image format (rim format) is developed for the usage in the similar-image retrieval. The format is based on PCA and ICA which can compress source images with an equivalent or often better rate-distortion than JPEG. Besides the data compression, the learned PCA/ICA bases are utilized in the similar-image retrieval since they reflect each source image’s local patterns. Following the format presentation, an image search viewer for network environments (Wisvi; Waseda image search viewer) is presented. Therein, each query is an image per se. The Wisvi system based on the “rim” method successfully finds similar images from non-uniform network environments. Experiments support that the PCA/ICA methods are viable to the joint compression and retrieval of digital images. Interested test users can download a β-version of the tool for the joint image compression and retrieval from a web site specified in this paper. TP-17 PID: 5001 Recognition of Proceeding Vehicles with Specific Information Hyo Jong Lee, Chonbuk National University, Korea Although the recognition of a license plate number or vehicle type has been researched, the recognition of vehicles using all features has not been studied due to its complexity. In this paper a novel method is proposed to identify vehicles with specific information, i.e., color, license plate, and vehicle’s model. Low level image processing and texture descriptors are computed from the front image of vehicles. Then, two three-layer neural networks were built and trained for license plate and vehicle’s model. TP-18 PID: 5043 Design Methodology of optimized IG_gHSOFPNN and Its application to pH neutralization process
TP
160
Ho-Sung Park, Wonkwang University, Korea Kyung-Won Jang, Wonkwang University, Korea Sung-Kwun Oh, The University of Suwon, Korea Tae-Chon Ahn, Wonkwang University, Korea In this paper, we propose design methodology of optimized Information granulation based genetically optimized Hybrid Self-Organizing Fuzzy Polynomial Neural Networks (IG_gHSOFPNN) by evolutionary optimization. The augmented IG_gHSOFPNN results in a structurally optimized structure and comes with a higher level of flexibility in comparison to the one we encounter in the conventional HSOFPNN. The GA-based design procedure being applied at each layer of IG_gHSOFPNN leads to the selection of preferred nodes (FPNs or PNs) available within the HSOFPNN. The obtained results demonstrate superiority of the proposed networks over the existing fuzzy and neural models. TP-19 PID: 5061 Pattern Classification using a Set of Compact Hyperspheres Amir Atiya, Cairo University, Egypt Sherif Hashem, Cairo University, Egypt Hatem Fayed, Cairo University, Egypt Prototype classifiers are one of the simplest and most intuitive approaches in pattern classification. However, they need careful positioning of prototypes to capture the distribution of each class region. Classical methods, such as learning vector quantization (LVQ), are sensitive to the initial choice of the number and the locations of the prototypes. To alleviate this problem, a new method is proposed that represents each class region by a set of compact hyperspheres. The number of hyperspheres and their locations are determined by setting up the problem as a set of quadratic optimization problems. Experimental results show that the proposed approach significantly beats LVQ and Restricted Coulomb Energy (RCE) in most performance aspects. TP-20 PID: 5104 A Neuro-Fuzzy-Based Agent System Using Data Distribution among the Agents Laura Emmanuella O Santana, Federal University of Rio Grande do, Brasil Anne M P Canuto, Federal University of Rio Grande do, Brasil Marjory C C Abreu, Federal University of Rio Grande do, Brasil The NeurAge system is a neuro-based multi-agent system and has been proposed as a result of the use of intelligent agents in the structure of multi-classifier systems. This system has presented good results in some conventional (centralized) and distributed classification tasks. In this paper, instead of neural networks, the NeurAge agents will be composed of neuro-fuzzy networks. The main aim of this paper is to investigate the benefits of using fuzzy concepts in a neuro-based multi-agent system and that data are distributed among its agents. TP-21 PID: 5105 SuperResolution Image Reconstruction Using a Hybrid Bayesian Approach
TP
161
Tao Wang , Zhengzhou Institute of Surveying and Mapping, China Yan Zhang , Zhengzhou Institute of Surveying and Mapping, China Yong Sheng Zhang, Zhengzhou Institute of Surveying and Mapping, China There are increasing demands for high-resolution (HR) images in various applications. Image super-resolution (SR) reconstruction refers to methods that increase image spatial resolution by fusing information from either a sequence of temporal adjacent images or multi-source images from different sensors. In the paper we propose a hybrid Bayesian method for image reconstruction, which firstly estimates the unknown point spread function (PSF) and an approximation for the original ideal image, and then sets up the HMRF image prior model and assesses its tuning parameter using maximum likelihood estimator, finally computes the regularized solution automatically. Hybrid Bayesian estimates computed on actual satellite images and video sequence show dramatic visual and quantitative improvements in comparison with the bilinear interpolation result, the projection onto convex sets (POCS) estimate and Maximum A Posteriori (MAP) estimate. TP-22 PID: 5186 A Fast and Efficient Route Finding Method for Car Navigation Systems with Neural Networks Mehdi Hashemzadeh, Islamic Azad University, Iran Mohammad. R. Meybodi, Amirkabir University of Technology, Iran In this paper we have proposed a new route finding method for car navigation systems meanwhile, have utilized learning power and high speed of neural networks in our proposed method. Our suggested procedures miss all limitations on standard algorithms planned for all these systems. In this way, despite of optimal rout finding between the origin and destination the parameters have been adopted than regarding roads traffics conditions, their limits and coincidence of routes in which all driver requests can simply be carried out. We believe that, this has been the first time discussing about neural networks on car navigation systems. Therefore, it could be the startup for widespread investigations. TP-23 PID: 5206 Applying an Intelligent Neural System to Predicting Lot Output Time in a Semiconductor Fabrication Factory Toly Chen, Feng Chia University, Taiwan, R.O.C. Output time prediction is a critical task to a wafer fab (fabrication plant). To further enhance the accuracy of wafer lot output time prediction, the concept of input classification is applied to Chen’s fuzzy back propagation network (FBPN) approach in this study by pre-classifying input examples with the k-means (kM) classifier before they are fed into the FBPN. Production simulation is also applied in this study to generate test examples. According to experimental results, the prediction accuracy of the intelligent neural system was significantly better than those of four existing approaches: BPN, case-based reasoning (CBR), FBPN without example classification, and evolving fuzzy rules (EFR), in most cases by achieving a 11%~46% (and an average of 31%) reduction in the root-mean-squared-error (RMSE) over the comparison basis – BPN.”
TP
162
TP-24 PID: 5214 Ultra-fast fMRI Imaging with high-fidelity activation map Neelam Sinha, Indian Institute of Science, India Manojkumar Saranathan, Indian Institute of Science, India A. G. Ramakrishnan, Indian Institute of Science, India Juan Zhou, NTU, Singapore Jagath C. Rajapakse, NTU, Singapore Functional Magnetic Resonance Imaging (fMRI) requires ultra-fast imaging in order to capture the on-going spatio-temporal dynamics of the cognitive task. We make use of correlations in both k-space and time, and thereby reconstruct the time series by acquiring only a fraction of the data, using an improved form of the well-known dynamic imaging technique k-t BLAST (Broad-use Linear Acquisition Speed-up Technique). k-t BLAST (k-tB) works by unwrapping the aliased Fourier conjugate space of k-t ( y-f space). The unwrapping process makes use of an estimate of the true y-f space, obtained by acquiring a blurred unaliased version. In this paper, we propose two changes to the existing algorithm. Firstly, we improve the map estimate using generalized series reconstruction. The second change is to incorporate phase constraints from the training map. The proposed technique is compared with existing k-tB on visual stimulation fMRI data obtained on 5 volunteers. Results show that the proposed changes lead to gain in temporal resolution by as much as a factor of 6. Performance evaluation is carried out by comparing activation maps obtained using reconstructed images, against that obtained from the true images. We observe up to 10dB improvement in PSNR of activation maps. Besides, RMSE reduction on fMRI images, of about 10% averaged over the entire time series, with a peak improvement of 35% compared to the existing k-tB, averaged over 5 data sets, is also observed. TP-25 PID: 5219 Speech Recognition with Multi-Modal Features Based on Neural Networks Myung Won Kim, Soongsil University, Korea Joung Woo Ryu, Electronics and Telecommunications Research Institute, Korea Eun Ju Kim, Soongsil University, Korea Recent researches have been focusing on fusion of audio and visual features for reliable speech recognition in noisy environments. In this paper, we propose a neural network based model of robust speech recognition by integrating audio, visual, and contextual information. Bimodal Neural Network (BMNN) is a multi-layer perceptron of 4 layers, which combines audio and visual features of speech to compensate loss of audio information caused by noise. In order to improve the accuracy of speech recognition in noisy environments, we also propose a post-processing based on contextual information which are sequential patterns of words spoken by a user. Our experimental results show that our model outperforms any single mode models. Particularly, when we use the contextual information, we can obtain over 90% recognition accuracy even in noisy environments, which is a significant improvement com-pared with the state of art in speech recognition. TP-26 PID: 5220
TP
163
Neuro-fuzzy modeling and Fuzzy rule extraction applied to Conflict Management Thando Tettey, University of the Witwatersrand, South Africa Tshilidzi Marwala, University of the Witwatersrand, South Africa This paper outlines all the computational methods which have been applied to the conflict management. A survey of all the pertinent literature relating to conflict management is also presented. The paper then introduces the Takagi-Sugeno fuzzy model for the analysis of interstate conflict. It is found that using interstate variables as inputs, the Takagi-Sugeno fuzzy model is able to forecast conflict cases with an accuracy of 80.36%. Furthermore, it found that the fuzzy model offers high levels of transparency in the form of fuzzy rules. It is then shown how these rules can be translated in order to validate the fuzzy model. The Takagi-Sugeno model is found to be suitable for interstate modeling as it demonstrates good forecasting ability while offering a transparent interpretation of the modeled rules. TP-27 PID: 5225 Kernel Uncorrelated Discriminant Analysis for Radar Target Recognition: A Comparative Study Ling Wang, Xidian University, China Liefeng Bo, Xidian University, China Licheng Jiao, Xidian University, China Kernel fisher discriminant analysis (KFDA) has received extensive study in recent years as a dimensionality reduction technique. KFDA always encounters an intrinsic singularity of scatter matrices in the feature space, namely ‘small sample size’ (SSS) problem. Several novel methods have been proposed to cope with this problem. In this paper, kernel uncorrelated discriminant analysis (KUDA) is proposed, which not only can bear on the SSS problem but also extract uncorrelated features, a desirable property for many applications. And then, we have conducted a comparative study on the application of KUDA and other variants of KFDA in radar target recognition problem. The experimental results indicate the effectiveness of KUDA and illustrate the utility of KFDA on the problem. TP-28 PID: 5230 Mitigating Deception in Genetic Search through Suitable Coding S. K. Basu, Banaras Hindu University, India A. K. Bhatia, Banaras Hindu University, India Formation of hamming cliff hampers the progress of genetic algorithm in seemingly deceptive problems. We demonstrate through an analysis of neighbourhood search capabilities of the mutation operator in genetic algorithm that the problem can sometimes be overcome through proper genetic coding. Experiments have been conducted on a 4-bit deceptive function and the pure-integer programming problem. The integer-coded genetic algorithm performs better and requires less time than the binary-coded genetic algorithm in these problems. TP-29 PID: 5244
TP
164
rotation invariant face detection using convolutional neural networks Fok Hing Chi Tivive, University of Wollongong, Australia Abdesselam Bouzerdoum, University of Wollongong, Australia This article addresses the problem of rotation invariant face detection using convolutional neural networks. Recently, we developed a new class of convolutional neural networks for visual pattern recognition. These networks have a simple network architecture and use shunting inhibitory neurons as the basic computing elements for feature extraction. Three networks with different connection schemes have been developed for in-plane rotation invariant face detection: fully-connected, toeplitz-connected, and binary-connected networks. The three networks are trained using a variant of Levenberg-Marquardt algorithm and tested on a set of 40,000 rotated face patterns. As a face/non-face classifier, these networks achieve 97.3% classification accuracy for a rotation angle in the range ±900 and 95.9% for full in-plane rotation. The proposed networks have fewer free parameters and better generalization ability than the feedforward neural networks, and outperform the conventional convolutional neural networks. TP-30 PID: 5338 An Automotive Detector Using Biologically Motivated Selective Attention Model for a Blind Spot Monitor Jaekyoung Moon, Kyungpook National University, Korea Jiyoung Yeo, Kyungpook National University, Korea Sungmoon Jeong, Kyungpook National University, Korea PalJoo Yoon, Mando Corporation Central R&D Center, Korea Minho Lee, Kyungpook National University, Korea The conventional side-view and rear-view mirrors are not enough for driver’s safety in an automobile. A driver may not be able to recognize the vehicle in a blind spot. In this paper, we propose an automotive detector algorithm using biologically motivated selective attention model for a blind spot monitor. This method decides a region of interest (ROI) which includes the blind spot from the successive image frames obtained by side-view cameras. It can detect the dangerous situations in the ROI using novelty points from the biologically motivated selective attention model, and alerts the driver whether there is dangerous object for changing the lane in driving. The proposed algorithm is based on deciding the ROI using difference from intensity histogram of a Gaussian smoothed image and finding the novelty points from the biologically motivated selective attention model. From variations of those novelty points, we determine whether a vehicle is approaching or not. TP-31 PID: 5348 Image Registration with Regularized Neural Network Anbang Xu, Beijing Normal University, China Ping Guo, Beijing Normal University, China Hua Li, Chinese Academy of Sciences, China In this paper, we propose a new method to improve the image registration accuracy in feedforward neural networks (FNN) based scheme. In the proposed method, Bayesian regularization is applied to improve the generalization capability of the FNN. The
TP
165
features extracted from the image sets by kernel independent component analysis (KICA) technique are input vectors of regularized FNN. The outputs of the neural network are those translation, rotation and scaling parameters with respect to reference and observed image sets. Comparative experiments are performed between FNN with regularization and without regularization under various conditions. The results show that the proposed method is much improved not only at accuracy but also remarkably at robust to noise. TP-32 PID: 5359 Adaptive Color Space Switching Based Approach for Face Tracking Chuan-Yu Chang, National Yunlin University of Science & Technology, Taiwan Yung-Chin Tu, Kun Shan University, Taiwan Hong-Hao Chang, National Yunlin University of Science & Technology, Taiwan In this paper, a support vector machine (SVM) based adaptive color switching for human face tracking is proposed. The color space is switching to the most appropriate color space model (CSM) according to circumstance conditions adaptively. Recently, many face tracking algorithms used empirical skin color model to discriminate skin/non-skin regions. These skin color models not consider illumination variation and result in less capacity to model skin color distribution. In this work, four color spaces and Laws texture extracted from face image database are used to train each SVM independently. In the pre-processing, the discrete wavelet transform (DWT) refines the face features would concentrate important features and reduce the computational complexity. Then, the features are transformed into four CSMs for SVMs which provide good generalization through optimal hyperplane. In testing, we perform quality measurement method to evaluate the face tracking performance and aggregating each SVM classification results to color space switching. Experimental results show that the proposed method would switch to the most appropriate color space according to quality measurement, automatically. TP-33 PID: 5364 Joint De-Interleaving/Recognition System of Radar Pulses Based on SVC and K-Means Qiang Guo, Harbin Engineering University,China Wanhai Chen, Harbin Engineering University,China Xingzhou Zhang, Harbin Engineering University,China Zheng Li, Southwest Research Institute of Electronic Equipment, China Di Guan, Harbin Normal University, China As radar signal environments become denser and radar signals become more complex, the task of an ESM operator becomes more difficult. This paper presented a de-interleaving/recognition system of radar pulses based on the combination of SVC and K-means clustering. Compared the conventional de-interleaving system, it can produce more complex and compact clustering boundaries according to the distribution characteristics of data set and has good generalization performance. The simulation experiment result shows that the system can sort efficiently radar signals in the high density and complex pulses environment. TP-34 PID: 5422 Training RBF Neural Network via Quantum-behaved Particle Swarm Optimization
TP
166
Jun Sun, Southern Yangtze University, China Wenbo Xu, Southern Yangtze University, China Jing Liu, Southern Yangtze University, China Radial Basis Function (RBF) networks are widely applied in function approximation, system identification, chaotic time series forecasting, etc. To use a RBF network, a training algorithm is absolutely necessary for determining the network parameters. The existing training algorithms, such as Orthogonal Least Squares (OLS) algorithm, clustering and gradient descent algorithm, have their own shortcomings. In this paper, we make an attempt to explore the applicability of Quantum-behaved Particle Swarm Optimization, a newly proposed evolutionary search technique, in training RBF neural network. The proposed QPSO-Trained RBF network was test on nonlinear system identification problem, and the results show that it can identifying the system more quickly and precisely than that trained by Particle Swarm algorithm. TP-35 PID: 5461 Neuro-Genetic Approach for Solving Constrained Nonlinear Optimization Problems Fabiana Cristina Bertoni, University of Sao Paulo, Brazil Ivan Nunes da Silva, University of Sao Paulo, Brazil This paper presents a neuro-genetic approach for solving constrained nonlinear optimization problems. Genetic algorithm must its popularity to make possible cover nonlinear and extensive search spaces. On the other hand, artificial neural networks have high computational rates due to the use of a massive number of simple processing elements and the high degree of connectivity between these elements. Neural networks with feedback connections provide a computing model capable of solving a large class of optimization problems. The association of a modified Hopfield network with genetic algorithm guarantees the convergence of the system to the equilibrium points, which represent feasible solutions for constrained nonlinear optimization problems. Simulated examples are presented to demonstrate that proposed method provides a significant improvement. TP-36 PID: 5531 Speech Feature Extraction Based on Wavelet Modulation Scale for Robust Speech Recognition Xin Ma, Shandong University, China Weidong Zhou, Shandong University, China Fang Ju, Shandong University, China Qi Jiang, Shandong University, China An analysis based on wavelet modulation scales feature extraction is proposed. Considering human auditory perception and varieties of disturbances, instead of the frequency differences, wavelet modulation scales are adopted to reflect the dynamic features of speech in ASR. Experiments for the Chinese digit-string recognition show extracting the wavelet modulation scales as the dynamic features have good performance both in additional noises and convolutional noises environment. TP-37
TP
167
PID: 5564 Manufacturing Yield Improvement by Clustering M. A. Karim, University of Melbourne, Australia S. Halgamuge, University of Melbourne, Australia A. J. R. Smith, University of Melbourne, Australia A. L. Hsu, University of Melbourne, Australia Dealing with product yield and quality in manufacturing industries is getting more difficult due to the increasing volume and complexity of data and quicker time to market expectations. Data mining offers tools for quick discovery of relationships, patterns and knowledge in large databases. Growing self-organizing map (GSOM) is established as an efficient unsupervised data-mining algorithm. In this study some modifications to the original GSOM are proposed for manufacturing yield improvement by clustering. These modifications include introduction of a clustering quality measure to evaluate the performance of the programme in separating good and faulty products and a filtering index to reduce noise from the dataset. Results show that the proposed method is able to effectively differentiate good and faulty products. It will help engineers construct the knowledge base to predict product quality automatically from collected data and provide insights for yield improvement. TP-38 PID: 5574 Applying neural network models to color emotion of product design Chung-Hsing Yeh, Monash University, Australia Yang-Cheng Lin, National Hualien University of Education, Taiwan This paper presents an approach applying a number of neural network (NN) models to color emotion of product design. We use 33 mobile phones with 50 colors individually to exam how a color affect a product image perceived by a consumer. Thus, we conduct an experimental study on mobile phones due to its various form and color. This paper also demonstrates the advantage of using neural networks for determining the optimal combination of product form and product color, particularly if the product form is already decided. This paper provides useful insights to save any amount of money and time for the new product development. TP-39 PID: 5589 Emotion-Based Classification of Movie Clips using Neural Networks Saowaluk C. Watanapa, King Mongkut’s University of Technology Thonburi, Thailand Bundit Thipakorn, King Mongkut’s University of Technology Thonburi, Thailand Nipon Charoenkitkarn, King Mongkut’s University of Technology Thonburi, Thailand Jonathan H. Chan, King Mongkut’s University of Technology Thonburi, Thailand This paper classifies movie clips into emotion-based categories using an artificial neural network. Using the concept of the film aesthetic theories, selected features of the low-level video and audio data that characterize semantic-level contents of movie clips were used as input to the neural network. The emotive classes of Excitement, Joy, and Sadness are chosen to be the output for this work. The results show that our model can classify the movie clips satisfactorily, with an 87.5% average accuracy.
TP
168
TP-40 PID: 4281 Forecasting Electricity Demand by Hybrid Machine Learning Model Shu Fan, Osaka Sangyo University, Japan Chengxiong Mao, Huazhong University of Science and Technology, China Jiadong Zhang, Osaka Sangyo University, Japan Luonan Chen, Osaka Sangyo University, Japan This paper proposes a hybrid machine learning model for electricity demand forecasting, based on Bayesian Clustering by Dynamics (BCD) and Support Vector Machine (SVM). In the proposed model, a BCD classifier is firstly applied to cluster the input data set into several subsets by the dynamics of load series in an unsupervised manner, and then, groups of 24 SVMs for the next day’s electricity demand curve are used to fit the training data of each subset. In the numerical experiment, the proposed model has been trained and tested on the data of the historical load from New York City.
FA401
169
Friday Session A, 6 October 2006
FA401 APNNA Presidential Invited Talks Section Chair: Manwai MAK, Hong Kong Polytechnic University Invited talk by Yixin Zhong Invited talk by Laiwan Chan
FA402
170
FA402 Special Session K-New Trends in Self-Organizing Maps II Section Chair: Tetsuo FURUKAWA, Kyushu Institute of Technology FA402-1 PID: 5590 An Online Adaptation Control System using mnSOM Shuhei Nishida, The University of Kitakyushu, Japan; Kazuo Ishii, Kyushu Institute of Technology, Japan; Tetsuo Furukawa, Kyushu Institute of Technology, Japan; Autonomous Underwater Vehicles (AUVs) are attractive tools to survey earth science and oceanography, however, there exists a lot of problems to be solved such as motion control, acquisition of sensor data, decision-making, navigation without collision, self-localization and so on. In order to realize useful and practical robots, underwater vehicles should take their action by judging the changing condition from their own sensors and actuators, and are desirable to make their behavior, because of features caused by the working environment. We have been investigated the application of brain-inspired technologies such as Neural Networks (NNs) and Self-Organizing Map (SOM) into AUVs. A new controller system for AUVs using Modular Network SOM (mnSOM) proposed by Tokunaga et al. is discussed in this paper. The proposed system is developed using recurrent NN type mnSOM. The efficiency of the system is investigated through the simulations. FA402-2 PID: 5583 Self-Organizing Map with Input Data Represented as Graph Takeshi Yamakawa, Kyushu Institute of Technology, Japan; Keiichi Horio, Kyushu Institute of Technology, Japan; Masaharu Hoshino, Ricoh Co., Ltd., Japan; This paper proposes a new method of Self-Organizing Map (SOM) in which an input space is represented as a graph by modifications of a distance measure and an updating rule. The distance between input node and reference element is defined by the shortest distance between them in the graph. The reference elements are updated along the shortest path to the input node. The effectiveness of the proposed method is verified by applying it to a Traveling Salesman Problem. FA402-3 PID: 5581 Language Learnability by Feedback Self-Organizing Maps Fuminori Mizushima, Kyushu Institute of Technology, Japan; Takashi Toyoshima, Kyushu Institute of Technology, Japan; Identification tasks are experimented on multi-layered selforganizing maps with feedback, using strings of category symbols of English sentences. Training is carried out in a setting that approximates language acquisition by human infants as closely as possible. Novel strings that are not used in training are tested whether they can be identified as grammatical or not. Test strings can be longer with no finite bounds than the trained strings, with recursions within themselves, just as natural language syntax allows.
FA402
171
FA402-4 PID: 5074 Evaluation-Based Topology Representing Network for Accurate Learning of Self-Organizing Relationship Network Takeshi Yamakawa, Kyushu Institute of Technology, Japan; Keiichi Horio, Kyushu Institute of Technology, Japan; Takahiro Tanaka, FANUC LTD, Japan; A Self-Organizing Relationship (SOR) network approximates a desirable input-output (I/O) relationship of a target system using I/O vector pairs and their evaluations. However, in the case where the topology of the network is different from that of the data set, the SOR network cannot precisely represent the topology of the data set and generate desirable outputs, because topology of the SOR network is fixed in one- or two-dimensional surface during learning. On the other hand, a Topology Representing Network (TRN) precisely represents the topology of the data set by a graph using the Competitive Hebbian Learning. In this paper, we propose a novel method which represents topology of the data set with evaluation by creating a fusion of SOR network and TRN. FA402-5 PID: 5256 A Digital Hardware Architecture of Self-Organizing Relationship (SOR) Network Hakaru Tamukoh, Kyushu Institute of Technology, Japan; Keiichi Horio, Kyushu Institute of Technology, Japan; Takeshi Yamakawa, Kyushu Institute of Technology, Japan; This paper describes a new algorithm of self-organizing relationship (SOR) network for an efficient digital hardware implementation and also presents its digital hardware architecture. In SOR network, the weighted average of fuzzy inference takes heavy calculation cost. To cope with this problem, we propose a fast calculation algorithm for the weighted average using only active units. We also propose a new generating technique of membership function by representing its width on power-of-two, which suits well with the digital hardware bit-shift process. The proposed algorithm is implemented on FPGA with massively parallel architecture. The experimental result shows that the proposed SOR network architecture has a good approximation ability of nonlinear functions.
FA403
172
FA403 Signal Processing Section Chair: Shouyuan YANG, Jiangxi University of Finance & Economics FA403-1 PID: 4859 Fuzzy Controllers Based QoS Routing Algorithm with a Multiclass Scheme for Ad Hoc Networks Chao Gui, Hubei University of Economics; Baolin Sun, Hubei University of Economics, Wuhan University of Technology, Wuhan University of Science and Engineering; As multimedia and group-oriented computing becomes increasingly popular for the users of mobile ad hoc networks (MANET). Due to the dynamic nature of the network topology and restricted resources, quality of service (QoS) and multicast routing in MANET is a challenging task. It attracts the interests of many people. In this paper, we present a fuzzy controllers based QoS routing algorithm with a multiclass scheme (FQRA) in MANET. The performance of this scheduler is studied using NS2 and evaluated in terms of quantitative metrics such as path success ratio, average end-to-end delay and throughput. Simulation shows that the approach is efficient, promising and applicable in MANET. FA403-2 PID: 5120 Direction of arrival Estimation based on minor component analysis approach Donghai Li, Zhengzhou Information Science and Technology Institute, China; Shihai Gao, Zhengzhou Information Science and Technology Institute, China; Feng Wang, Zhengzhou Information Science and Technology Institute, China; Fankun Meng, Zhengzhou Information Science and Technology Institute, China; Many high resolution DOA estimation algorithms like MUSIC and ESPRIT estimation are based on the sub-space concept and require the eigen-decomposition of the input correlation matrix. As quantities of computation of eigen-decomposition, it is unsuitable for real time processing. An algorithm for noise subspace estimation based on minor component analysis is proposed. These algorithms are based on anti-Hebbian learning neural network and contain only relatively simple operations, which are stable, convergent, and have self-organizing properties. Finally a method of real-time parallel processing is proposed, and data processing can be finished at end time of sampling. Simulations show that the proposed algorithm has an analogy performance with the MUSIC algorithm. FA403-3 PID: 5568 Two-Stage Temporally Correlated Source Extraction Algorithm with Its Application in Extraction of Event-Related Potentials Zhi-Lin Zhang, Shanghai Jiao Tong University, China; Liqing Zhang, Shanghai Jiao Tong University, China; Xiu-Ling Wu, Shanghai Jiao Tong University, China; Jie Li, Shanghai Jiao Tong University, China; Qibin Zhao, Shanghai Jiao Tong University, China;
FA403
173
To extract source signals with certain temporal structures, such as periodicity, we propose a two-stage extraction algorithm. Its first stage uses the autocorrelation property of the desired source signal, and the second stage exploits the independence assumption. The algorithm is suitable to extract periodic or quasi-periodic source signals, without requiring that they have distinct periods. It outperforms many existing algorithms in many aspects, confirmed by simulations. Finally, we use the proposed algorithm to extract the components of visual event-related potentials evoked by three geometrical figure stimuli, and the classification accuracy based on the extracted components achieves 93.2%. FA403-4 PID: 5178 An Adaptive Beamforming by A Generalized Unstructured Neural Network Askin Demirko, University of Missouri-Rolla, USA; Levent Acar, University of Missouri-Rolla, USA; Robert S. Woodley, 21st Century Systems, Inc., USA; In this paper, an adaptive array beamforming by an unstructured neural network based on the mathematics of holographic storage is presented. This work is inspired by similarities between brain waves and the wave propagation and subsequent interference patterns seen in holograms. Then the mathematics to produce a general mathematical description of the holographic process is analyzed. From this analysis it is shown that how the holographic process can be used as an associative memory network. Additionally, the process may also be used a regular feed-forward network. The most striking aspect of these networks is that, using the holographic process, the apriori knowledge of the system may be better utilized to tailor the neural network for an adaptive beamforming problem. This aspect, makes this neural network formation process particularly useful for the beamforming. FA403-5 PID: 4826 Learning Method for Signal Reconstruction from Samples with Jitter error Shouyuan Yang, Naikai University, China; Xingwei Zhou, Naikai University, China; Zhanjie Song, Tianjin University, China; The main purpose of this paper is to investigate the reconstruction of signals from its sampling data with random jitter errors. Inspired by Smale and Zhou[10, 10], we use the regularized learning scheme to reconstruct signals and images from the sampling data with random jitter errors. Furthermore, the specific error estimations are given. These results may be used in digital communication, simulation and image processing.
FA404
174
FA404 Special Session G-Incremental Self-Organizing Networks Section Chair: Saman HALGAMUGE, Arthur HSU, University of Melbourne; Damminda ALAHAKOON, Monash University FA404-1 PID: 5538 Scalable Dynamic Self-Organising Maps for Mining Massive Textual Data Yu Zheng Zhai, University of Melbourne, Australia; Arthur Hsu, University of Melbourne, Australia; Saman K Halgamuge, University of Melbourne, Australia; Traditional text clustering methods require enormous computing resources, which make them inappropriate for processing large scale data collections. In this paper we present a clustering method based on the word category map approach using a two-level Growing Self-Organising Map (GSOM). A significant part of the clustering task is divided into separate subtasks that can be executed on different computers using the emergent Grid technology. Thus enabling the rapid analysis of information gathered globally. The performance of the proposed method is comparable to the traditional approaches while improves the execution time by 15 times. FA404-2 PID: 5537 Semi-Supervised Learning of Dynamic Self-Organising Maps Arthur Hsu, University of Melbourne, Australia; Saman K. Halgamuge, University of Melbourne, Australia; We present a semi-supervised learning method for the Growing Self- Organising Maps (GSOM) that allows fast visualisation of data class structure on the 2D network. Instead of discarding data with missing values, the network can be trained from data with up to 60% of their class labels and 25% of attribute values missing, while able to make class prediction with over 90% accuracy for the benchmark datasets used. FA404-3 PID: 4546 Self-organizing by information maximization Ryotaro Kamimura, Tokai University, Japan; The present paper shows that a self-organizing process can be realized simply by maximizing information between input patterns and competitive units. We have already shown that information maximization corresponds to competitive processes. Thus, if cooperation processes can be incorporated in information maximization, self-organizing maps can naturally be realized by information maximization. By using the weighted sum of distances among neurons or collected distance, we successfully incorporate cooperation processes in the main mechanism of information maximization. For comparing our method with the standard SOM, we applied the method to the well-known artificial data and show that clear feature maps can be obtained by maximizing information. FA404-4
FA404
175
PID: 5275 A new SOM algorithm for electricity load forecasting Manuel Mart´ın-Merino, Universidad Pontificia de Salamanca, Spain; Jesus Rom´an, Universidad Pontificia de Salamanca, Spain; The interest in electricity load forecasting has grown up in the last years. However, the accurate load prediction remains a difficult task due particularly to the non-linear character of the time series and the periodical and seasonal patterns it exhibits. Several machine learning techniques such as the Support Vector Machines (SVM) have been developed that are able to deal with nonlinear time series. However, the patterns of electricity demand change strongly and periodically with seasons, holidays and other factors. Therefore global models such as the SVM are not expected to perform well. In this paper we propose a new segmentation algorithm based on the Self Organizing Maps (SOM) to split the time series into homogeneous regions. Next, a linear SVM is locally trained in each region. The algorithm proposed has been applied to the prediction of the maximum daily electricity demand. The experimental results show that the new segmentation algorithm helps to improve several well known forecasting techniques. FA404-5 PID: 4490 ART-based parallel learning of growing SOMs and its application to TSP Tetsunari Oshime, Hosei University, Japan; Toshimichi Saito, Hosei University, Japan; Hiroyuki Torikai, Hosei University, Japan; This paper studies parallel learning of growing self-organizing maps ( GSOMs ) and its application to traveling sales person problems ( TSPs ). Input space of city positions are divided into subspaces automatically through adaptive resonance theory ( ART ) map. One GSOM is allocated to each subspace and grows following input data. After all the GSOMs grow sufficiently they are fused and we obtain a tour. The algorithm performance can be controlled by four parameters: the number of subspaces, insertion interval, learning coefficient and final number of cells. In basic experiments for a data-set of 929 cities we can find semi-optimal solution much faster than serial methods although there exists trade-off between tour length and execution time. FA404-6 PID: 4427 Smooth Seamless Surface Construction Based on Conformal Self-organizing Map Cheng-Yuan Liou, National Taiwan University, Taiwan; Yen-Ting Kuo, National Taiwan University, Taiwan; Jau-Chi Huang, National Taiwan University, Taiwan; This paper presents a method to construct a smooth seamless conformal surface for the genus-0 manifold. The method is developed for the conformal self-organizing map [10]. The constructed surface is both piecewise smooth and continuous. The mapping between the model surface and the sphere surface is one-to-one and onto. We show experiments in surface reconstruction and texture mapping.
FA405
176
FA405 Special Session F - Advances in Machine Learning Methods based Pattern Recognition Section Chair: Kaizhu HUANG, Fujitsu China FA405-1 PID: 5546 Investigation on Multisets Mixture Learning Based Object Detection Zhi-Yong Liu, Chinese Academy of Sciences, China; Hong Qiao, Chinese Academy of Sciences, China; Lei Xu, Chinese University of Hong Kong, Hong Kong; By minimizing the mean square reconstruction error, multisets mixture learning (MML) provides a general approach for object detection in image. To calculate each sample reconstruction error, as the object template is represented by a set of contour points, the MML needs inefficiently enumerate the distances between the sample and all the contour points. In this paper we develop the line segment approximation (LSA) algorithm to calculate the reconstruction error, which is shown theoretically and experimentally more efficient than the enumeration method. It is also experimentally illustrated that the MML based algorithm has a better noise resistance ability than the generalized Hough transform (GHT) based counterpart. FA405-2 PID: 5560 Projective Non-negative Matrix Factorization with Applications to Facial Image Processing Zhirong Yang, Helsinki University of Technology, Finland; Zhijian Yuan, Helsinki University of Technology, Finland; Jorma Laaksonen, Helsinki University of Technology, Finland; We propose a new variant of Non-negative Matrix Factorization (NMF), including its model and two optimization rules. Our method is based on positively constrained projections and is related to the conventional SVD or PCA decomposition. The new model can potentially be applied to image compression and feature extraction problems. Of the latter, we consider processing of facial images, where each image consists of several parts and for each part the observations with different lighting mainly distribute along a straight line through the origin. No regularization terms are required in the objective functions and both suggested optimization rules can easily be implemented by matrix manipulations. The experiments show that the derived base vectors are more spatially localized than those of NMF. In turn, the better part-based representations improve the recognition rate of semantic classes such as the gender or existence of mustache in the facial images. FA405-3 PID: 5562 Modular RBF Neural Networks for Texture Classification Chuan-Yu Chang, National Yunlin University of Science & Technology, Taiwan; Shih-Yu Fu, National Yunlin University of Science & Technology, Taiwan;
FA405
177
Image classification had become an important technique in multimedia processing. Recently, neural network-based methods had been proposed to resolve the classification problem. Among them, the radial basis function (RBF) neural network is the most popular architecture because it has good learning and approximate capacity. Since the traditional RBF neural networks were sensitive to center initialization. To obtain appropriate centers, it needs to find the significant features for further RBF clustering. In addition, the training procedure of the traditional RBF is time-consuming. Therefore, in this paper, a combination of self-organizing map (SOM) neural network and learning vector quantization (LVQ) is proposed to select more appropriate centers for RBF network, and a modular RBF (MRBF) neural network is proposed to improve the classification rate and speed up the training time. The experimental results show that the proposed MRBF network has better performance than the discrete wavelet frame (DWF) and the rotated wavelet filter (RWF) methods in image classification. FA405-4 PID: 5540 A Novel Sports Video Logo Detector Based on Motion Analysis Hongliang Bai, Chinese Academy of Sciences, China; Wei Hu, Intel China Research Center, China; Tao Wang, Intel China Research Center, China; Xiaofeng Tong, Intel China Research Center, China; Changping Liu, Chinese Academy of Sciences, China; Yimin Zhang， Intel China Research Center, China; Replays are key cues for events detection in sport videos since they are the immediate consequence of highlights or important events happened in sports. In many sports videos, replays are usually sandwiched with two identical logo transitions, prompt the beginning and end of a replay. A logo transition is a kind of special digital video effects, usually contains 12-35 consecutive frames, describe a flying or variable object. In this paper, a novel automatic logo detection approach is proposed. It contains two main stages: a logo transition template is automatically learned by dynamic programming and unsupervised clustering, a key frame is also extracted; then the extracted key frame and the learned logo template are used jointly to detect logos in sports videos. The optical flow features are used to depict the motion characteristics of the logo transitions. Experiments on different types of sports videos show that the proposed approach can reliably detect logos in sports videos efficiently. FA405-5 PID: 5239 A Hybrid Handwritten Chinese Address Recognition Approach Kaizhu Huang, Fujitsu R&D Center Ltd, China; Jun Sun, Fujitsu R&D Center Ltd, China; Yoshinobu Hotta, Fujitsu Laboratories Ltd, Japan; Katsuhito Fujimoto, Fujitsu Laboratories Ltd, Japan; Satoshi Naoi,Fujitsu Laboratories Ltd, Japan; Chong Long, Tsinghua University, China; Li Zhuang, Tsinghua University, China; Xiaoyan Zhu, Tsinghua University, China;
FA405
178
Handwritten Chinese Address Recognition describes a difficult yet important pattern recognition task. There are three difficulties in this problem: (1) Handwritten address is often of free styles and of high variations, resulting in inevitable segmentation errors. (2) The number of Chinese characters is large, leading low recognition rate for single Chinese characters. (3) Chinese address is usually irregular, i.e., different persons may write the same address in different formats. In this paper, we propose a comprehensive and hybrid approach for solving all these three difficulties. Aiming to solve (1) and (2), we adopt an enhanced holistic scheme to recognize the whole image of words (defined as a place name) instead of that of single characters. This facilitates the usage of address knowledge and avoids the difficult single character segmentation problem as well. In order to attack (3), we propose a hybrid approach that combines the word-based language model and the holistic word matching scheme. Therefore, it can deal with various irregular addresses. We provide theoretical justifications, outline the detailed steps, and perform a series of experiments. The experimental results on various real addresses demonstrate the advantages of our novel approach.
FA406
179
FA406 Data & Text Processing Section Chair: Manuel MARTIN-MERINO, Universidad Pontificia de Salamanca FA406-1 PID: 5397 Stimulus Related Data Analysis by Structured Neural Networks Bernd Br¨uckner, Leibniz Institute for Neurobiology, Germany; In the analysis of biological data artificial neural networks are a useful alternative to conventional statistical methods. Because of its advantage in analyzing time courses the Multilevel Hypermap Architecture (MHA) is used for analysis of stimulus related data, exemplified by fMRI studies with auditory stimuli. Results from investigations with the MHA show an improvement of discrimination in comparison to statistical methods. With an interface to the well known BrainVoyager software and with a GUI for MATLAB an easy usability of the MHA and a good visualization of the results are possible. The MHA is an extension of the Hypermap introduced by Kohonen. By means of the MHA it is possible to analyze structured or hierarchical data (data with priorities, data with context, time series, data with varying exactness), which is difficult or impossible to do with known self-organizing maps so far. FA406-2 PID: 4588 Visualization of depending Patterns in Metabonomics Stefan Roeder, UFZ – Centre for Environmental Research Leipzig-Halle Ltd., Germany; Ulrike Rolle-Kampczyk, UFZ – Centre for Environmental Research Leipzig-Halle Ltd., Germany; Olf Herbarth, UFZ – Centre for Environmental Research Leipzig-Halle Ltd., Germany; This paper describes an approach for visualization of patterns in large data sets. The data sets are combined from external exposure and internal stress factors on human health. For deduction of modes of action on human health, external and internal stress factors have to be combined and classified. The approach shown in this paper is based upon clustering algorithms. Relationships between cases ban be obtained by visual inspection of clustering results. FA406-3 PID: 5543 Knowledge as Basis Broker〞The Research of Matching Customers Problems and Professionals M?tiers Ruey-Ming Chao, Day-Yeh University, Taiwan; Chi-Shun Wang, Day-Yeh University, Taiwan; With the popularization of the concept knowledge economic management, it not only propels the whole development of knowledge economy but also directs the industry of “Basic Agent Service” of becoming the mainstream in the present markets. This research institute constructed a system called, “Knowledge-based Broker Service Center” (KBSC). It allows the customers to submit economic or business field questions online in the form of their natural language. By using Chinese phrase-cutting, key words weighted value calculations, and professional categorizations, it can automatically analyze the nature of
FA406
180
the customer’s problem and search for the relevant information in the HR database to list the most suitable names of specialists as the assigned coordinator for the clients. When each matching procedure was finished, the questionnaire was given to examine the correctness of the data search following adjustments of the system. FA406-4 PID: 5263 A local semi-supervised Sammon algorithm for textual data visualization Manuel Mart´ın-Merino, Universidad Pontificia de Salamanca, Spain; Alberto Mu˜noz, Universidad Carlos III de Madrid, Spain; Sammon’s mapping is a powerful non-linear technique that allows us to visualize high dimensional object relationships. It has been applied to a broad range of practical problems and particularly to the visualization of the semantic relations among terms in textual databases. However, the word maps generated by the Sammon mapping suffer from a low discriminant power due to its unsupervised nature and to the “curse of dimensionality”. Fortunately several search engines such as Yahoo provides a manually created classification of a subset of documents that may help to overcome this problem. In this paper we propose a semi-supervised Sammon algorithm that takes advantage of this classification to improve the discriminant power of the word maps generated. The algorithm has been tested using a real textual collection and evaluated through several objective functions. The experimental results suggest that the new model improves significantly well known unsupervised alternatives.
FA407
181
FA407 Computer Vision Section Chair: Usman KHAN, Brunel University FA407-1 PID: 5450 Camera pose estimation by an artificial neural network Ryan G. Benton, The University of Louisiana at Lafayette, United States of America; Chee-hung Henry Chu, The University of Louisiana at Lafayette, United States of America; Reconstruction of a three-dimensional scene using images taken from two views is possible if the relative pose of the cameras is known. A traditional approach to estimating the pose of the cameras uses eight pairs of corresponding points and involves the solution of a set of homogeneous equations. We propose a multi-layered feedforward network solution. Empirical results demonstrate the feasibility of using the network to recover the relative pose of the cameras in the three- dimensional world. FA407-2 PID: 4788 Implicit Camera Calibration By Using Resilient Neural Networks P?nar C?ivicio?glu, Erciyes University; Erkan Be?sdok, Erciyes University; The accuracy of 3D measurements of objects is highly affected by the errors originated from camera calibration. Therefore, camera calibration has been one of the most challenging research fields in the computer vision and photogrammetry recently. In this paper, an Artificial Neural Network Based Camera Calibration Method, NBM, is proposed. The NBM is especially useful for back-projection in the applications that do not require internal and external camera calibration parameters in addition to the expert knowledge. The NBM offers solutions to various camera calibration problems such as calibrating cameras with automated active lenses that are often encountered in computer vision applications. The difference of the NBM from the other artificial neural network based back-projection algorithms used in intelligent photogrammetry (photogrammetron) is its ability to support the multiple view geometry. In this paper, a comparison of the proposed method has been made with the Bundle Block Adjustment based back-projection algorithm, BBA. The performance of accuracy and validity of the NBM have been tested and verified over real images by extensive simulations. FA407-3 PID: 5049 Implicit Camera Calibration Using an Artificial Neural Network Dong-Chul Park, Myong Ji University, Korea; Dong-Min Woo, Myong Ji University, Korea; A camera calibration method based on a nonlinear modeling function of an artificial neural network (ANN) is proposed in this paper. With the application of the nonlinear mapping feature of an ANN, the proposed method successfully finds the relationship between image coordinates without explicitly calculating all the camera parameters, including position, orientation, focal length, and lens distortion. Experiments on the
FA407
182
estimation of 2-D coordinates of image world given 3-D space coordinates are performed. In comparison with Tsai’s two stage method, the proposed method reduced modeling errors by 11.45% on average. FA407-4 PID: 5445 3D Freeform Surfaces from Planar Sketches using Neural Networks Usman Khan, Brunel University, United Kingdom; Abdelaziz Terchi, Brunel University, United Kingdom; Sungwoo Lim, Brunel University, United Kingdom; David Wright, Brunel University, United Kingdom; Sheng-Feng Qin, Brunel University, United Kingdom; A novel intelligent approach into 3D freeform surface reconstruction from planar sketches is proposed. A multilayer perceptron (MLP) neural network is employed to induce 3D freeform surfaces from planar freehand curves. Planar curves were used to represent the boundaries of a freeform surface patch. The curves were varied iteratively and sampled to produce training data to train and test the neural network. The obtained results demonstrate that the network successfully learned the inverse-projection map and correctly inferred the respective surfaces from fresh curves. FA407-5 PID: 5189 General Adaptive Transfer Functions Design for Volume Rendering By Using Neural Networks Liansheng Wang, National University of Defense Technology, China; Xucan Chen, National University of Defense Technology, China; Sikun Li, National University of Defense Technology, China; Xun Cai, National University of Defense Technology, China; In volume data visualization, the classification is used to determine voxel visibility and is usually carried out by transfer functions that define a mapping between voxel value and color/opacity. The design of transfer functions is a key process in volume visualization applications. However, one transfer function that is suitable for a data set usually dose not suit others, so it is difficult and time-consuming for users to design new proper transfer function when the types of the studied data sets are changed. By introducing neural networks into the transfer function design, a general adaptive transfer function (GATF) is proposed in this paper. Experimental results showed that by using neural networks to guide the transfer function design, the robustness of volume rendering is promoted and the corresponding classification process is optimized. FA407-6 PID: 4691 Real-time Synthesis of 3D Animations by Learning Self-Organizing Mixture Networks of Parametric Gaussians Yi WANG, Tsinghua University,China; Lei XIE, City University of Hong Kong, Hong Kong; Zhi-Qiang LIU, City University of Hong Kong, Hong Kong; Li-Zhu ZHOU, Tsinghua University, China;
FA407
183
In this paper, we present a novel real-time approach to synthesize 3D character animations with required style by adjusting a few parameters or scratching mouse cursor. The productivity of our approach comes from learning captured 3D human motion as a self-organizing mixture network (SOMN) of parametric Gaussians. The learned model describes motions under control of a vector variable called style variable, and acts as a probabilistic mapping from the low-dimensional style values to the high-dimensional 3D poses. We designed a pose synthesis algorithm and developed an easy-to-use graphical interface program to allow the users, especially the animators, to generate poses by given a style values. We designed an interesting method called style-interpolation, which accepts a sparse sequence of key style values and interpolates a dense sequence of style values to synthesize a segment of animation. This keystyling method is able to produce animations that are more realistic and natural-looking than those synthesized with the traditional keykeyframing technique.
FA408
184
FA408 Theoretical Modeling and Analysis IV Section Chair: Andrew LEUNG, City University of Hong Kong FA408-1 PID: 4584 Free Energy of Stochastic Context Free Grammar on Variational Bayes Tikara Hosino, Tokyo Institute of Technology, Japan; Kazuho Watanabe, Tokyo Institute of Technology, Japan; Sumio Watanabe, Tokyo Institute of Technology, Japan; Variational Bayesian learning is proposed for approximation method of Bayesian learning. In spite of efficiency and experimental good performance, their mathematical property has not yet been clarified. In this paper we analyze variational Bayesian Stochastic Context Free Grammar which includes the true distribution thus the model is nonidentifiable. We derive their asymptotic free energy. It is shown that in some prior conditions, the free energy is much smaller than identifiable models and satisfies eliminating redundant non-terminals. FA408-2 PID: 4485 Asymptotic Behavior of Stochastic Complexity of Complete Bipartite Graph-type Boltzmann Machines Yu Nishiyama, Tokyo Institute of Technology, Japan; Sumio Watanabe, Tokyo Institute of Technology, Japan; In singular statistical models, it was shown that Bayes learning is effective. However, on Bayes learning, calculation containing the Bayes posterior distribution requires huge computational costs. To overcome the problem, mean field approximation (or equally variational Bayes method) was proposed. Recently, the generalization error and stochastic complexity in mean field approximation have been theoretically studied. In this paper, we treat the complete bipartite graph-type Boltzmann machines and derive the upper bound of the asymptotic stochastic complexity in mean field approximation. FA408-3 PID: 5462 Improving VG-RAM neural networks performance using knowledge correlation Raphael V. Carneiro, Universidade Federal do Espírito Santo, Brasil; Stiven S. Dias, Universidade Federal do Espírito Santo, Brasil; Dijalma Fardin Jr., Universidade Federal do Espírito Santo, Brasil; Hallysson Oliveira, Universidade Federal do Espírito Santo, Brasil; Artur S. d’Avila Garcez, City University, London, United Kingdom; Alberto F. De Souza, Universidade Federal do Espírito Santo, Brasil; In this work, the correlation between input-output patterns stored in the memory of the neurons of Virtual Generalizing RAM (VG-RAM) weightless neural networks, or knowledge correlation, is used to improve the performance of these neural networks. The knowledge correlation, detected using genetic algorithms, is used for changing the distance function employed by VGRAM neurons in their recall mechanism. In order to evaluate the performance of the method, experiments with several well-known datasets
FA408
185
were made. The results showed that VG-RAM networks employing knowledge correlation perform significantly better than standard VG-RAM networks. FA408-4 PID: 5332 The Bifurcating Neuron Network 3 As Coloring Problem Solver and N-ary Associative Memory Jinhyuk Choi, Information and Communications University, Korea; Geehyuk Lee, Information and Communications University, Korea; The Bifurcating Neuron (BN) is an integrate-and-fire neuron that exhibits crisis-mediated transitions between multiple, symmetrical chaotic attractors. The Bifurcating Neuron Network 3 (BNN-3), a class of BN networks, was reported to be a natural model for solving coloring problems due to the multi-stability of BN. An important question left behind unanswered by the preliminary report was the scalability of BNN-3 as a coloring problem solver. Another question was the possibility of BNN-3 playing an N-ary associative memory as other multi-state neuron network models do. We carried out an extended study and were able to reach positive conclusions for both questions. FA408-5 PID: 5346 Prediction error of a fault tolerant neural network John Sum, Chung Shan Medical University, Taiwan; Chi-sing Leung, City University of Hong Kong, Hong Kong; Kevin Ho, Providence University, Taiwan; For more than a decade, prediction error has been one powerful tool to measure the performance of a neural network. In this paper, we extend the technique to a kind of fault tolerant neural network. Consider a neural network to be suffering from multiple-node fault, a formulae similar to that of Generalized Prediction Error has been derived. Hence, the effective number of parameter of such a fault tolerant neural network is obtained. A difficulty in obtaining the mean prediction error is discussed and then a simple procedure for estimation of the prediction error empirically is suggested. FA408-6 PID: 5412 Heat Diffusion Classifier on a Graph Haixuan Yang, The Chinese University of Hong Kong, Hong Kong; Michael R. Lyu, The Chinese University of Hong Kong, Hong Kong; A classifier is recently proposed based on a heat diffusion model on a finite K nearest neighbor (KNN) graph. To broaden the scope of the previous work, we separate the graph construction procedure and classifier construction procedure by proposing a Heat Diffusion Classier on a Graph (G-HDC). This separation allows various graph inputs. Besides traditional KNN graph, we propose two other candidate graph inputs for G-HDC: a graph (SKNN) with the shortest edges whose number is the same as the KNN graph, and the Minimum Spanning Tree (MST). These two graphs can be also considered as the representation of the underlying geometry, and the heat diffusion model on them can be considered as the approximation to the way that heat flows on a geometric structure. KNN graph, SKNN graph, and MST lead to three classifiers when applied to the G-HDC:
FA408
186
KNN-HDC, SKNN-HDC, and MST-HDC. Experiments show that SKNN-HDC and MST-HDC can be considered as two other candidate classifiers, which enrich the family of heat diffusion classifiers.
FB401
187
Friday Session B, 6 October 2006
FB401 Invited Talk by Aapo Hyvarinen Section Chair: Laiwan CHAN, The Chinese University of Hong Kong
FB403
188
FB403 Data Processing & Information Security Section Chair: Takayuki KIMURA, Saitama University FB403-1 PID: 4981 The Design of Data-link Equipment Redundant Strategy Li Qian-Mu, Nanjing University, China; Xu Man-Wu, Nanjing University, China; Zhang Hong, Nanjing University of Science and Technology, China; Liu Feng-Yu, Nanjing University of Science and Technology, China; A framework model proposed in this paper is a data-link Equipment Redundant Strategy based on reliability theory. The strategy combined with the normal maintenance could greatly improve the performance of the network system. The static-checking and policy of authentication mechanism ensure the running network without any error. The redundant equipments are independent but are capable of communication with each other when they work their actions. The model is independent of specific application environment, thus providing a general-purpose framework for fault diagnosis. An example is given to express the calculating method. FB403-2 PID: 4893 Improved Real-time Intrusion Detection System Byung-Joo Kim, Youngsan University, Korea; Il Kon Kim, Kyungpook National University, Korea; We developed earlier version of real-time intrusion detection system using empirical kernel map combining least squares SVM(LS- SVM). It consists of two parts. One part is feature extraction by empirical kernel map and the other one is classification by LS-SVM. The main problem of earlier system is that it is not operated real-time because LS- SVM is executed in batch way. In this paper we propose an improved real time intrusion detection system incorporating earlier developed system with incremental LS-SVM. Applying the proposed system to KDD CUP 99 data, experimental results show that it has a remarkable feature extraction, classification performance and reducing detection time compared to earlier version of real-time intrusion detection system. FB403-3 PID: 5255 A new algorithm for packet routing problems using chaotic neurodynamics Takayuki Kimura, Saitama University, Japan; Tohru Ikeguchi, Saitama University, Japan; We propose a new algorithm for packet routing problems using chaotic neurodynamics. We first compose a basic neural network which routes packets using information of shortest path lengths from a node to the other nodes in the computer network. When the computer network topology is regular, the basic routing method works well, however, when the computer network topology becomes irregular, it doesn’t work well. The reason is that most of packets cannot be transmitted to their destinations because of packet congestion in the computer network. To avoid such an undesirable problem, we extended
FB403
189
the basic method to employ chaotic neurodynamics. We confirm that our proposed method exhibits good performance for scale-free networks. Furthermore, we analyze why the proposed routing method is effective, introducing the method of surrogate data which is often used in the field of nonlinear time-series analysis. Using such a statistical control, we confirm that using chaotic neurodynamics is the most effective policy to decentralize the packet congestion in the computer network.
FB404
190
FB404 Pattern Classification & Recognition V Section Chair: Eric YU, ASTRI FB404-1 PID: 4301 Bark Classification Based on Gabor Filter Features Using RBPNN Neural Network Zhi-Kai Huang, Chinese Academy of Sciences, China; De-Shuang Huang, Chinese Academy of Sciences, China; Ji-Xiang Du, Chinese Academy of Sciences, China; Zhong-Hua Quan, Chinese Academy of Sciences, China; Shen-Bo Guo, Chinese Academy of Sciences, China; This paper proposed a new method of extracting texture features based on Gabor wavelet. In addition, the application of these features for bark classification applying radial basis probabilistic network (RBPNN) has been introduced. In this method, the bark texture feature is firstly extracted by filtering the image with different orientations and scales filters, then the mean and standard deviation of the image output are computed, the image which have been filtered in the frequency domain. Finally, the obtained Gabor feature vectors are fed up into RBPNN for classification. Experimental results show that, first, features extracted using the proposed approach can be used for bark texture classification. Second, compared with radial basis function neural network (RBFNN), the RBPNN achieves higher recognition rate and better classification efficiency when the feature vectors have low-dimensions. FB404-2 PID: 4892 A Morphological Neural Network Approach for Vehicles Detection from High Resolution Satellite Imagery Hong Zheng, Wuhan University, China; Li Pan, Wuhan University, China; Li Li, Wuhan University, China; This paper introduces a morphological neural network approach to extract vehicle targets from high resolution panchromatic satellite imagery. In the approach, the morphological shared-weight neural network (MSNN) is used to classify image pixels on roads into vehicle targets and non-vehicle targets, and a morphological preprocessing algorithm is developed to identify candidate vehicle pixels. Experiments on 0.6 meter resolution QuickBird panchromatic data are reported in this paper. The experimental results show that the MSNN has a good detection performance. FB404-3 PID: 5002 Secure Personnel Authentication Based on Multi- modal Biometrics under Ubiquitous Environments Dae-Jong Lee, Chungbuk National University, Korea; Man-Jun Kwon, Chungbuk National University, Korea; Myung-Geun Chun, Chungbuk National University, Korea;
FB404
191
In this paper, we propose a secure authentication method based on multimodal biometrics system under ubiquitous computing environments. For this, the face and signature images are acquired in PDA and then each image with user ID and name is transmitted via WLAN (Wireless LAN) to the server and finally the PDA receives authentication result from the server. In the proposed system, face recognition algorithm is designed by PCA and LDA. On the other hand, the signature verification is designed by a novel method based on grid partition, Kernel PCA and LDA. To calculate the similarity between test image and training image, we adopt the selective distance measure determined by various experiments. More specifically, Mahalanobis and Euclidian distance measures are used for face and signature, respectively. As the fusion step, decision rule by weighted sum fusion scheme effectively combines the two matching scores calculated in each biometric system. From the real-time experiments, we convinced that the proposed system makes it possible to improve the security as well as user’s convenience under ubiquitous computing environments.
FB405
192
FB405 Biomedical Applications III Section Chair: Chaojian SHI, Shanghai Maritime University FB405-1 PID: 5487 Gene Feature Extraction using T-Test Statistics and Kernel Partial Least Squares Shutao Li, Hunan University, China; Chen Liao, Hunan University, China; James Kwok, Hong Kong University of Science and Technology, Hong Kong; In this paper, we propose a gene extraction method by using two standard feature extraction methods, namely the T-test method and kernel partial least squares (KPLS), in tandem. First, a preprocessing step based on the T-test method is used to filter irrelevant and noisy genes. KPLS is then used to extract features with high information content. Finally, the extracted features are fed into a classifier. Experiments are performed on three benchmark datasets: breast cancer, ALL/AML leukemia and colon cancer. While using either the T-test method or KPLS does not yield satisfactory results, experimental results demonstrate that using these two together can significantly boost classification accuracy, and this simple combination can obtain state-of-the-art performance on all three datasets. FB405-2 PID: 5179 Connectionist Approaches for Predicting Mouse Gene Function from Gene Expression Emad Andrews Shenouda, University of Toronto, Canada; Quaid Morris, University of Toronto, Canada; Anthony J. Bonner, University of Toronto, Canada; Identifying gene function has many useful applications especially in Gene Therapy. Identifying gene function based on gene expression data is much easier in prokaryotes than eukaryotes due to the relatively simple structure of prokaryotes. That is why tissue-specific expression is the primary tool for identifying gene function in eukaryotes. However, recent studies have shown that there is a strong learnable correlation between gene function and gene expression. This paper outlines a new approach for gene function prediction in mouse. The prediction mechanism depends on using Artificial Neural Networks (NN) to predict gene function based on quantitative analysis of gene co-expression. Our results show that neural networks can be extremely useful in this area. Also, we explore clustering of gene functions as a preprocessing step for predicting gene function. FB405-3 PID: 5091 Wavelet Spectral Entropy for Indication of Epileptic Seizure in Extracranial EEG Xiaoli Li, The University of Birmingham, UK; Automated detection of epileptic seizures is very important for an EEG monitoring system. In this paper, a continuous wavelet transform is proposed to calculate the spectrum of scalp EEG data, the entropy and a scale-averaged wavelet power are extracted to indicate the epileptic seizures by using a moving window technique. The
FB405
193
tests of five patients with different seizure types show wavelet spectral entropy and scale-averaged wavelet power are more efficiency than renormalized entropy and Kullback_Leiler (K-L) relative entropy to indicate the epileptic seizures. We suggest that the measures of wavelet spectral entropy and scale-averaged wavelet power should be contained to indicate the epileptic seizures in a new EEG monitoring system. The abstract should summarize the contents of the paper and should contain at least 70 and at most 150 words. It should be set in 9-point font size and should be inset 1.0 cm from the right and left margins. There should be two blank (10-point) lines before and after the abstract. This document is in the required format. FB405-4 PID: 5385 Automatic Detection of Critical Epochs in coma-EEG using ICA and High Order Statistics G. Inuso, DIMET - Mediterranea University of Reggio Calabria, Italy; F. La Foresta, DIMET - Mediterranea University of Reggio Calabria, Italy; N. Mammone, DIMET - Mediterranea University of Reggio Calabria, Italy; F.C. Morabito, DIMET - Mediterranea University of Reggio Calabria, Italy; Previous works showed that the joint use of Principal Component Analysis (PCA) and Independent Component Analysis (ICA) allows to extract a few meaningful dominant components from the EEG of patients in coma. A procedure for automatic critical epoch detection might support the doctor in the long time monitoring of the patients, this is why we are headed to find a procedure able to automatically quantify how much an epoch is critical or not. In this paper we propose a procedure based on the extraction of some features from the dominant components: the entropy and the kurtosis. This feature analysis allowed us to detect some epochs that are likely to be critical and that are worth inspecting by the expert in order to assess the possible restarting of the brain activity. FB405-5 PID: 5504 A Hybrid Algorithm of Snake Energy Model and Particle Swarm Optimization to Infer Genetic Networks Cheng-Long Chuang, National Taiwan University, Taiwan; Chung-Ming Chen, National Taiwan University, Taiwan; Grace S. Shieh, Academia Sinica, Taiwan; A pattern recognition approach, based on shape feature extraction, is proposed to infer genetic networks from time course microarray data. The proposed algorithm learns patterns from known genetic interactions, such as RT-PCR confirmed gene pairs, and tunes the parameters using particle swarm optimization algorithm. This work also incorporates a score function to separate significant predictions from non-significant ones. The prediction accuracy of the proposed method applied to data sets in Spellman et al. (1998) is as high as 91%, and true-positive rate and false-negative rate are about 61% and 1%, respectively. Therefore, the proposed algorithm may be useful for inferring genetic interactions.
FB406
194
FB406 Data & Text Processing Section Chair: Chenghua LI, Chonbuk National University FB406-1 PID: 5140 Ontology learning from text: A soft computing paradigm Rowena Chau, Monash University, Australia; Kate Smith-Miles, School of Engineering and Information Technology, Australia; Chung-Hsing Yeh, Monash University, Australia; Text-based information accounts for more than 80% of today’s Web content. They consist of Web pages written in different natural languages. As the semantic Web aims at turning the current Web into a machine-understandable knowledge repository, availability of multilingual ontology thus becomes an issue at the core of a multilingual semantic Web. However, multilingual ontology is too complex and resource intensive to be constructed manually. In this paper, we propose a three-layer model built on top of a soft computing framework to automatically acquire a multilingual ontology from domain specific parallel texts. The objective is to enable semantic smart information access regardless of language over the Semantic Web. FB406-2 PID: 4627 Text Categorization Based on Artificial Neural Networks Cheng Hua Li, Chonbuk National University, Korea; Soon Choel Park, Chonbuk National University, Korea; This paper described two kinds of neural networks for text categorization, multi-output perceptron learning (MOPL) and back-propagation neural network (BPNN), and then we proposed a novel algorithm using improved back-propagation neural network. This algorithm can overcome some shortcomings in traditional back-propagation neural network such as slow training speed and easy to enter into local minimum. We compared the training time and the performance, and tested the three methods on the standard Reuter-21578. The results show that the proposed algorithm is able to achieve high categorization effectiveness as measured by the precision, recall and F-measure. FB406-3 PID: 5536 Word Frequency Effect and Word similarity Effect in Korean Lexical Decision Task and Their Computational Model YouAn Kwon, Korea University, Korea; KiNam Park, Korea University, Korea; HeuiSeok Lim, Hanshin University, Korea; KiChun Nam, Korea University, Korea; Soonyoung Jung, Korea University, Korea; In this paper, we investigate whether the word frequency effect and the word similarity effect could be applied to Korean lexical decision task (henceforth, LDT). Also we propose a computational model of Korean LDT and present comparison results between human and computational model on Korean LDT. We found that the word frequency
FB406
195
effect and the similarity effect in Korean LDT were language general phenomena in both the behavioral experiment and the proposed computational simulation.
FB407
196
FB407 Manufacturing Applications Section Chair: Shuichi KUROGI, Kyusu Institute of Technology FB407-1 PID: 4544 Application of ICA in On-line Verification of the Phase Difference of the Current Sensor Xiaoyan Ma, Chinese Academy of Sciences, China; Huaxiang Lu, Chinese Academy of Sciences, China; The performance of the current sensor in power equipment may become worse affected by the environment. In this paper, based on ICA, we propose a method for on-line verification of the phase difference of the current sensor. However, not all source components are mutually independent in our application. In order to get an exact result, we have proposed a relative likelihood index to choose an optimal result from different runs. The index is based on the maximum likelihood evaluation theory and the independent subspace analysis. The feasibility of our method has been confirmed by experimental results. FB407-2 PID: 4816 Gear Crack Detection Using Kernel Function Approximation Weihua Li, South China University of Technology; Tielin Shi, Huazhong University of Science and Technology; Kang Ding, South China University of Technology; Failure detection in machine condition monitoring involves a classification mainly on the basis of data from normal operation, which is essentially a problem of one-class classification. Inspired by the successful application of KFA (Kernel Function Approximation) in classification problems, an approach of KFA-based normal condition domain description is proposed for outlier detection. By selecting the feature samples of normal condition, the boundary of normal condition can be determined. The outside of this normal domain is considered as the field of outlier. Experiment results indicated that this method can be effectively and successfully applied to gear crack diagnosis. FB407-3 PID: 5205 Ensemble of Competitive Associative Nets for Stable Learning Performance in Temperature Control of RCA Cleaning Solutions Shuichi Kurogi, Kyushu Institute of Technology, Kitakyushu, Japan; Daisuke Kuwahara, Kyushu Institute of Technology, Kitakyushu, Japan; Hiroaki Tomisaki, Kyushu Institute of Technology, Kitakyushu, Japan; Takeshi Nishida, Kyushu Institute of Technology, Kitakyushu, Japan; Mitsuru Mimata, Komatsu Electronics Inc., Japan; Katsuyoshi Itoh, Komatsu Electronics Inc., Japan; For cleaning silicon wafers via the RCA clean, temperature control is important in order to obtain a stable performance, but it is difficult mainly because the RCA solutions expose nonlinear and time-varying exothermic chemical reactions. So far, the MSPC (model switching predictive controller) using the CAN2 (competitive associative net 2)
FB407
197
has been developed and the effectiveness has been validated. However, we have observed that the control performance, such as overshoot and settling time, does not always improve as the number of learning iterations increases when using multiple units of the CAN2. So we apply the ensemble learning scheme to the CAN2 for stable control over learning iterations, and we examine the effectiveness of the present method by means of computer simulation. FB407-4 PID: 5574 Applying neural network models to color emotion of product design Chung-Hsing Yeh, Monash University, Australia; Yang-Cheng Lin, National Hualien University of Education, Taiwan; This paper presents an approach applying a number of neural network (NN) models to color emotion of product design. We use 33 mobile phones with 50 colors individually to exam how a color affect a product image perceived by a consumer. Thus, we conduct an experimental study on mobile phones due to its various form and color. This paper also demonstrates the advantage of using neural networks for determining the optimal combination of product form and product color, particularly if the product form is already decided. This paper provides useful insights to save any amount of money and time for the new product development.
FB408
198
FB408 Theoretical Modeling and Analysis V Section Chair: Haixuan YANG, The Chinese University of Hong Kong FB408-1 PID: 5103 Self-organized Path Constraint Neural Network: Structure and Algorithm Hengqing Tong, Wuhan University of Technology, China; Li Xiong, Wuhan University of Technology, China; Hui Peng, Wuhan University of Technology, China; Due to its flexibility and self-determination, self-organized learning neural network(NN) has been widely applied in many fields. Meanwhile, it has a well trend to develop. In our research, we find that structural equation modeling (SEM) may be reconstructed into a self-organized learning neural network, but the algorithm of NN need to be improved. In this paper, we first present an improved partial least square (PLS) algorithm in SEM using a suitable iterative initial value with constraint of unit vector. Then we propose a new self-organized path constraint neural network(SPCNN) based on SEM. Furthermore, we give the topology structure of SPCNN, describe the learning algorithm of SPCNN, including common algorithm and algorithm with a suitable initial weights value, and elaborate the function of SPCNN. FB408-2 PID: 4397 Training RBFs Networks: A Comparison among Supervised and not Supervised Algorithms Joaqu´ın Torres-Sospedra, Universitat Jaume I, Spain; Carlos Hern´andez-Espinosa, Universitat Jaume I, Spain; Mercedes Fern´andez-Redondo, Universitat Jaume I, Spain; In this paper, we present experiments comparing different training algorithms for Radial Basis Functions (RBF) neural networks. In particular we compare the classical training which consist of an unsupervised training of centers followed by a supervised training of the weights at the output, with the full supervised training by gradient descent proposed recently in same papers. We conclude that a fully supervised training performs generally better. We also compare Batch training with Online training and we conclude that Online training suppose a reduction in the number of iterations. FB408-3 PID: 4690 Soft Analyzer Modeling for Dearomatisation Unit Using KPCR with Online Eigenspace Decomposition Haiqing Wang, Zhejiang University, China; Daoying Pi, Zhejiang University, China; Ning Jiang, Zhejiang university of Technology, China; Steven X. Ding, University of Duisburg-Essen, Germany; The application of kernel method to petrochemical industry is explored in this paper. A nonlinear soft analyzer for the flashpoint measurement of Dearomatization process is developed by using kernel principal component regression (KPCR) method. To trace the
FB408
199
time varying dynamics and reject disturbances, a novel online eigenspace decomposing algorithm is proposed to update that of the Kernel Matrix, which is much faster than direct decomposition and meanwhile has stable numerical performance. Simulation results indicate the developed soft analyzer has satisfying prediction precision under both nominal and faulty operating conditions. FB408-4 PID: 5319 A New Approach to Discover Interlacing Data Structures in High-dimensional Space Tao Ban, Kobe University, Japan; Shigeo Abe, Kobe University, Japan; Changshui Zhang, Tsinghua University, China; Zhongbao Kou, Tsinghua University, China; Wei Shu, Tsinghua University, China; The discovery of structures hidden in high-dimensional data space is of great significance for understanding and further processing of the data. Real world datasets are often composed of bunches of low dimensional patterns, the interlacement of which may impede our ability to grasp the global structure of the data. Few of the existing methods focus on the detection and extraction of the manifolds representing distinct patterns. Inspired by the nonlinear dimensionality reduction method ISOMAP, we present a novel approach called Multi-Manifold Partition (MMP) to identify the interlacing low dimensional patterns. The algorithm has three steps: first a neighborhood graph is built to capture the intrinsic topological structure of the input data, then the dimensional uniformity of neighboring nodes is analyzed to discover the segments of patterns, finally the segments which are possibly from the same low-dimensional global structure are combined to obtain a global representation. Experiments on synthetic data as well as real problems are reported. The results show that this new approach to exploratory data analysis is effective and may boost our understanding of the data.
FP
200
Friday Poster Session III, 6 October 2006
FP-1 PID: 4473 Development of the Neuroinformatics Base Platform System: XooNIps Kazutsuna Yamaji, RIKEN Brain Science Institute, Japan Hiroyuki Sakai, RIKEN Brain Science Institute, Japan Yoshihiro Okumura, RIKEN Brain Science Institute, Japan Kazue Hosokawa, RIKEN Brain Science Institute, Japan Osamu Kurosaki, RIKEN Brain Science Institute, Japan Shiro Usui, RIKEN Brain Science Institute, Japan The developing field of neuroinformatics includes technologies for the collection and sharing of neuro-related digital resources. These resources will be of increasing value for understanding the brain. Developing a database system to integrate these disparate resources is necessary for full use of these resources. This study proposes a base database system termed XooNIps that utilizes the content management system called XOOPS. XooNIps is designed for developing databases in different research fields through customization of the option menu. According to the site policy, several types of user authorization, reviewing system, communicating function between the databases etc. are customized by GUI operations. XOOPS modules containing news, forums schedules, blogs and other information can be combined to enhance XooNIps functionality. These features offer better scalability, extensibility to the general neuroinformatics community. FP-2 PID: 4504 Content-based 3D Graphic Information Retrieval Soochan Hwang, Hankuk Aviation University, Korea Yonghwan Kim, Hankuk Aviation University, Korea Supporting the content-based retrieval of 3D graphic data has been little addressed in the most current 3D graphic systems. The systems focus on visualizing 3D images. This paper presents a 3D graphic data model which models 3D scenes using domain objects and their spatial relations. The model also supports the content-based query. The user can pose a visual query involving various 3D graphic features such as an inclusion of a given object, object’s shape, descriptive information, and spatial relations on the web interface. We discuss the 3D data modeling technique and content-based retrieval in detail. FP-3 PID: 4643 A Study on the configuration Control of a Mobile Manipulator Base upon the Optimal Cost Function Kwan-Houng Lee, Cheongju University, Korea Tae-jun Cho, Cheongju University, Korea In this paper, using the pre-determined specific tasks, a solid and complete solution for the optimal control of the mobile manipulator is proposed based on a divide and conquer scheme. In the scheme, a mobile manipulator is virtually divided into a mobile robot and a task robot. All the tasks are also divided into task segments that can be performed by
FP
201
only the task robot. An optimal configuration of the task robot is defined by the task oriented manipulability measure for given task segment. And using a cost function for optimality defined as a combination of the square errors of the desired and actual configurations of the mobile robot and of the task robot, the job which the mobile manipulator performs is optimized. We figured out the solution for the optimal configuration of a mobile manipulator with a series of tasks. FP-4 PID: 4658 Intrusion Detection by Backpropagation Neural Networks with Sample-Query and Attribute-Query Ray-I Chang, National Taiwan University, Taiwan Liang-Bin Lai, National Taiwan University, Taiwan Kevin Chan, Chung-Shan Institute of Science and Technology, Taiwan Sheng-Jyh Chen, Chung-Shan Institute of Science and Technology, Taiwan Jen-Shaing Kouh, National Taiwan University, Taiwan The growing network intrusions have put companies and organizations at a much greater risk of loss. In this paper, we propose a new learning methodology towards developing a novel intrusion detection system (IDS) by backpropagation neural networks (BPN) with sample-query and attribute-query. We test the proposed method by a benchmark intrusion dataset to verify its feasibility and effectiveness. Results show that choosing good attributes and samples will not only have impact on the performance, but also on the overall execution efficiency. The proposed method can significantly reduce the training time required. Additionally, the training results are good. It provides a powerful tool to help supervisors analyze, model and understand the complex attack behavior of electronic crime. FP-5 PID: 4678 Application of Neural-Network Inverse Dynamic online Learning Control on Physical Exoskeleton Heng Cao, East China University of Science and Technology, China Yuhai Yin, East China University of Science and Technology, China Ding Du, East China University of Science and Technology, China Jianjun Yan, East China University of Science and Technology, China Wenjin Gu, Navy Aeronautical Engineering College, China Exoskeleton system which is to assist the motion of physically weak persons such as disabled, injured and elderly persons is discussed in this paper. The proposed exoskeletons are controlled basically based on the electro-moyogram (EMG) signals. And a mind model is constructed to identify person’s mind for predicting or estimating person’s behavior. The proposed mind model is installed in an exoskeleton power assistive system named IAE for walking aid. The neural-network is also be used in this system to help learning. The on-line learning adjustment algorithm based on multi-sensor that are fixed on the robot is designed which makes the locomotion stable and adaptable. FP-6 PID: 4683
FP
202
Improved Clustering and Anisotropic Gradient Descent Algorithm for Compact RBF Network Delu Zeng, South China University of Technology, China Shengli Xie, South China University of Technology, China Zhiheng Zhou, South China University of Technology, China In the formulation of radial basis function (RBF) network, there are three factors mainly considered, i.e., centers, widths, and weights, which significantly affect the performance of the network. Within thus three factors, the placement of centers is proved theoretically and practically to be critical. In order to obtain a compact network, this paper presents an improved clustering (IC) scheme to obtain the location of the centers. What is more, since the location of the corresponding widths does affect the performance of the networks, a learning algorithms referred to as anisotropic gradient descent (AGD) method for designing the widths is presented as well. In the context of this paper, the conventional gradient descent method for learning the weights of the networks is combined with that of the widths to form an array of couple recursive equations. The implementation of the proposed algorithm shows that it is as efficient and practical as GGAP-RBF. FP-7 PID: 4791 Features Selection of SVM and ANN Using Particle Swarm Optimization for Power Transformers Incipient Fault Symptom Diagnosis Tsair-Fwu Lee, Chang Gung Memorial Hospital, Taiwan Fu-Min Fang, Chang Gung Memorial Hospital, Taiwan Ming-Yuan Cho, National Kaohsiung University of Applied Science, Taiwan For the purpose of incipient power transformer fault symptom diagnosis, a successful adaptation of the particle swarm optimization (PSO) algorithm to improve the performances of Artificial Neural Network (ANN) and Support Vector Machine (SVM) is presented in this paper. A PSO-based encoding technique is applied to improve the accuracy of classification, which removed redundant input features that may be confusing the classifier. Experiments using actual data demonstrated the effectiveness and high efficiency of the proposed approach, which makes operation faster and also increases the accuracy of the classification. FP-8 PID: 4862 New Region of Interest Image Coding Using Partial Bitplane Layered Shift for Medical Image Compression ZHANG Li-bao, Beijing Normal University, China Efficient image storage and transmission is significant for medical community. Regions of Interest (ROI) coding technique enables the regions important to medical diagnosis to be encoded and transmitted at higher quality than the other regions, which improves coding efficiency and reduces transmission time. In this paper, a new approach for ROI coding so-call partial bitplane layered shift (PBLShift) is presented. In this method, all bitplanes of image coefficients are composed of four parts-most significant ROIs bitplanes, most significant BG Bitplanes, general significant ROIs Bitplanes, least significant BG bitplanes. For single ROI, the new method can encode ROI for any
FP
203
quality by up-shift most significant ROIs bitplanes and down-shift least significant BG bitplanes. When multiple regions of interest are encoded, it can encode and transmit a certain number bitplanes of ROIs according to different degrees of interest. The presented method has lower coding complexity than the general scaling-based method and more flexibility than Maxshift for multiple ROI coding. Simulation experiments on the CT, MRI and state-of-the-art images show that PBLShift, in addition to alleviating the drawbacks of both ROI coding methods in JPEG2000, can support arbitrarily shaped multiple ROI coding with different degrees of interest without coding the ROI shapes. FP-9 PID: 4881 An architecture-adaptive neural network online control system Xun Liang, Peking University, China Jian Yang, Peking University, China Bin Lu, Peking University, China Rongchang Chen, National Taichung Institute of Technology, Taiwan In this paper, we present an architecture-adaptive intelligent self-tuning control system. The system incorporates a visual and user-friendly interface for neurocontrol. The system is composed of the supervisor module, the model refinement module, the process plant and the database. In the supervisor module, the user prescribes the desired curve for the plant dynamic process. The user does not have to be an expert in operations. The only knowledge she should possess is the best plant dynamic process according to the industry experience. This illustrative and intelligent way allows the operator to watch and manage more processes and focus on the key factors in control process. The novel point of this paper is that the desired best control curves, as opposed to the parameters, are designated by the operator, making the process more interactive and visualizing. The model refinement module is in parallel with the process plant and cost calculation model, and consists of the self-tuning process model, which contains an architecture-adaptive neural network. The model refinement module could learn intelligently the real process plant by the prompt adjustments based on the difference of the outputs of the two modules, and its learned model is also refined gradually. All the parameters and process data are saved into the database. This diagram is specially versatile in the complex nonlinear and time-variant systems in practice. FP-10 PID: 4894 A Neural Network based Software Retrieval System with Fuzzy-Related Thesaurus Huilin Ye, The University of Newcastle, Australia The qualities of both the classification and retrieval queries have significant impacts on the retrieval performance of a software retrieval system. A classification scheme based on a Nested Self-Organising Map (NSOM) and a query refinement based on a fuzzy-related thesaurus were proposed to promote the qualities. An NSOM consists of a top map and a set of nested maps. The retrieval on the top map maintains high recall while the retrieval on the nested maps enhances precision. A fuzzy-related thesaurus can be generated from an NSOM. The user can reformulate an improved query by adding terms or replacing an original query term with a related term stored in the thesaurus. The experimental results reveal that both the NSOM and query refinement significantly improved the retrieval performance.
FP
204
FP-11 PID: 4898 A DGC-based Data Classification Method Used for Abnormal Network Intrusion Detection Bo Yang, Wuhan University of Science and Technology, China Lizhi Peng, Jinan University, China Yuehui Chen, Jinan University, China Hanxing Liu, Wuhan University of Science and Technology, China Runzhang Yuan,Wuhan University of Science and Technology, China The data mining techniques used for extracting patterns that represent abnormal network behavior for intrusion detection is an important research area in network security. This paper introduces the concept of gravitation and gravitation field into data classification by utilizing analogical inference, and studied the method to calculate data gravitation. Based on the theoretical model of data gravitation and data gravitation field, the paper presented a new classification model called Data Gravitation based Classifier (DGC). The proposed approach was applied to an Intrusion Detection System (IDS) with 41 inputs (features). Experimental results show that the proposed method was efficient in data classification and suitable for abnormal detection using network processor-based platforms. FP-12 PID: 4902 Using Inverse Neural Network for HIV Adaptive Control B. Leke Betechuoh, University of Witwatersrand, South Africa T. Marwala, University of Witwatersrand, South Africa T. Tettey, University of Witwatersrand， South Africa Neural Networks are used in this paper, in an inverse configuration, for the adaptive control of HIV status of individuals. In the paper, a control mechanism to modify the status of individuals from positive to negative using the demographic properties (in this case, education) is implemented. Preliminary design showed inverse neural network to outperform the other methodology. The moral behind this implementation is to understand whether HIV can be controlled by modifying some of the demographic properties such as education. The proposed method is used on the HIV data set. It is found that the proposed method is able to predict the educational level of individuals to an accuracy of 88%, thus being possible to control the HIV status of the individuals based on educational level. FP-13 PID: 4903 Effectiveness of feature space selection on credit engineering on multi-group classification cases Junghee Park, Sogang University, Korea Kidong Lee, University of Incheon, Korea This study tests the sensitivity of input feature space selection on credit rating using four classifiers as backpropagation(BP), Kohonen selforganizing feature map, discriminant analysis(DA), and logistic regression. The results of the study are that BP network outperforms two statistical counterparts while Kohonen network shows the least accuracy
FP
205
among the models. The results also show that the selection of the feature spaces to the accuracy outcome may not be very sensitive. FP-14 PID: 4960 An Approach to Network Intrusion Detection Algorithm Based on BP-HMM Huang Guang-qiu, Xian University of Architecture & Technology, China Wang Xiao-hai, Xian University of Architecture & Technology, China A network intrusion detection framework and its associated algorithm based on BP-HMM are put forward; the training and recognition methods of the algorithm are given. A sheer classifier based on HMM can’t give attention to both the strong recognition ability for the corresponding objects and the maximizing difference lain in different models, so the BP neural network is used to provide state probability output for HMM in the HMM framework. Because of the coarse classification of BP, the limitation of HMM is overcome; the ability of classification and recognition is enhanced. Through the use of the any-path method, the accurate rate of recognition is not only improved, but also the obvious calculation predominance is obtained. FP-15 PID: 4972 RVM Ensemble for Text Classification Catarina Silva， School of Technology and Management of the Polytechnic Institute of Leiria, Portugal； Bernardete Ribeiro， University of Coimbra, Portugal； Automated classification of texts by their likeness or affinity has greatly eased the management and processing of the massive volumes of information we face everyday. Although Support Vector Machines (SVM) provide a state-of-the-art technique to tackle this problem, Relevance Vector Machines (RVM), which rely on Bayesian inference learning, offer advantages such as their capacity to find sparser and probabilistic solutions. A known problem with the Bayesian approaches, however, is their relative inability to scale to larger problems where millions of documents are involved as well as real-time user’s requests. We propose an ensemble strategy to circumvent RVMs scalability problem by applying a divide-and-conquer technique to handle the overload of available data, where the training documents are first divided amongst small RVM classifiers, then the ensemble combines their individual contributions. The solution achieved maintains a sparse decision function and is computationally efficient. Results with respect to REUTERS-21578 clearly demonstrate the proposed strategy is able to surpass other techniques, in terms of both classification performance and response time. FP-16 PID: 4976 Intelligent System for Features Extraction of Oil Slicks in SAR Images: Speckle Filters Analysis Danilo L. de Souza, Federal University of Rio Grande do Norte, Brazil Adri?ao D.D. Neto, Federal University of Rio Grande do Norte, Brazil Wilson da Mata, Federal University of Rio Grande do Norte, Brazil
FP
206
The development of automatic techniques for oil slick identification on the sea surface, captured through remote sensing images, cause a positive impact to a complete monitoring of the oceans and seas. C-band SAR (ERS-1, ERS-2, Radarsat and Envisat projects) is well adapted to detect ocean pollution because the backscatter is reduced by oil slick. This work propose a system for segmentation and feature extraction of oil slicks candidates based on techniques of digital image processing (filters, gradients, mathematical morphology) and artificial neural network (ANN). Different algorithms of speckle filtering are tested and a comparison for the considered system is presented. The process is thought to possess a level of automatization that minimizes the intervention of a human operator, being possible the processing of larger amount data. The focus of the work is to present a study detailed for feature extraction block proposed (architecture used and computational tools). FP-17 PID: 5009 3-D Reconstruction of Blood Vessels Skeleton Based on Neural Network Zhiguo Cao, Huazhong University of Sci&Tec, China Bo Peng, Huazhong University of Sci&Tec, China For the 3-D reconstruction of blood vessels skeleton from biplane angiography system, an efficient 3-D reconstruction method based on neural network(NN) is proposed in this paper. First, we find a set of 2-D corresponding points on the vessels’ skeleton in the matched image pair. Secondly, NN is utilized to build the relationships between the 3-D points and their projective 2-D points. Thirdly, by feeding the corresponding points we found into the NN, the 3D coordinates of the points on vessels can be obtained. At last we employ B-spline interpolation to improve reconstruction performance. Experiment results demonstrate the efficiency of the new method. FP-18 PID: 5036 Fault Diagnosis in Nonlinear Circuit Based on Volterra Series and Recurrent Neural Network Haiying Yuan, University of Electronic Science and Technology of China, China Guangju Chen, University of Electronic Science and Technology of China, China The neural network diagnosis method based on fault features denoted by frequency domain kernel in nonlinear circuit was presented here. Each order frequency domain kernel of circuit response under all fault states can be got by vandermonde method; the circuit features extracted was preprocessed and regarded as input samples of neural network, faults is classified. The uniform recurrent arithmetical formula of each order frequency-domain kernel was given, the Volterra frequency-domain kernel acquisition method was discussed, and the fault diagnosis method based on recurrent neural network was showed. A fault diagnosis illustration verified this method. The fault diagnosis method showed the advantages: no precise circuit model is needed in avoiding the difficulty in identifying nonlinear system online, less computation amount, high fault diagnosis efficiency. FP-19 PID: 5078
FP
207
A Distributed Neural Network Learning Algorithm For Network Intrusion Detection System Yanheng Liu, Jilin University, China Daxin Tian, Jilin University, China Xuegang Yu, Jilin University, China Jian Wang, Jilin University, China To make network intrusion detection systems can be used in Gigabit Ethernet, a distributed neural network learning algorithm (DNNL) is put for-ward to keep up with the increasing network throughput. The main idea of DNNL is splitting the overall traffic into subsets and several sensors learn them in parallel way. The advantage of this method is that the large data set can be split randomly thus reduce the complicacy of the splitting algorithm. The experiments are performed on the KDD’99 Data Set which is a standard intrusion detection benchmark. Comparisons with other approaches on the same bench-mark show that DNNL can perform detection with high detection rate. FP-20 PID: 5121 Evolving Hierarchical RBF Neural Networks for Breast Cancer Detection Yuehui Chen, Jinan University, China Yan Wang, Jinan University, China Bo Yang, Jinan University, China Hierarchical RBF networks consist of multiple RBF networks assembled in different level or cascade architecture. In this paper, an evolved hierarchical RBF network was employed to detect the breast cancel. For evolving a hierarchical RBF network model, Extended Com- pact Genetic Programming (ECGP), a tree-structure based evolutionary algorithm and the Differential Evolution (DE) are used to find an optimal detection model. The performance of proposed method was then compared with Flexible Neural Tree (FNT), Neural Network (NN), and RBF Neural Network (RBF-NN) by using the same breast cancer data set. Simulation results show that the obtained hierarchical RBF network model has a fewer number of variables with reduced number of input features and with the high detection accuracy. FP-21 PID: 5159 A Novel Blind Digital Watermark Algorithm Based on Neural Network and Chaotic Map Pengcheng Wei, Chongqing Education College, China Wei Zhang, Chongqing Education College, China Huaqian Yang, Chongqing Education College, China Degang Yang, Chongqing University, China In order to enhance robustness and security of the embedded watermark, proposed a novel blind digital watermark algorithm based on neural network and chaotic map. Firstly, a better chaotic sequence is generated by Cellular Neural Network (CNN) and Chebyschev map, using the chaotic sequence encrypted the watermark and its spectrum is spread. Then, BPN is trained to memorize the relationship among pixels of each sub-block image. Furthermore, the adaptive embedding algorithm is adopted to enhance the characters of the watermarking system. Simulation results are given which show that this scheme is practical, secure and robust.
FP
208
FP-22 PID: 5187 Intelligent Control System Based on Vague Neural Network Yibiao Zhao, Beijing Jiaotong University, China Siwei Luo, Beijing Jiaotong University, China Liang Wang, Beijing Jiaotong University, China Aidong Ma, Rui Fang, Beijing Jiaotong University, China Reinforcement learning is a class of model-free learning control method that can solve Markov decision problems. But it has some problems in applications, especially in MDPs of continuous state spaces. In this paper, based on the vague neural networks, we propose a Q-learning algorithm which is comprehensively considering the reward and punishment of the environment. Simulation results in cart-pole balancing problem illustrate the effectiveness of the proposed method. FP-23 PID: 5215 Performance Improvement in Collaborative Recommendation Using Multi-Layer Perceptron Myung Won Kim, Soongsil University, Korea Eun Ju Kim, Soongsil University, Korea Recommendation is to offer information which fits user’s interests and tastes to provide better services and to reduce information overload. It recently draws attention upon Internet users and information providers. Collaborative filtering is one of the widely used methods for recommendation. It recommends an item to a user based on the reference users’ preferences for the target item or the target user’s preferences for the reference items. In this paper, we propose a neural network based collaborative filtering method. Our method builds a model by learning correlation between users or items using a multi-layer perceptron. We also investigate integration of diverse information to solve the sparsity problem and selecting the reference users or items based on similarity to improve performance. We finally demonstrate that our method outperforms the existing methods through experiments using the EachMovie data. FP-24 PID: 5250 Automated Parameter Selection for Support Vector Machine Decision Tree Gyunghyun Choi, Hanyang University, Korea Suk Joo Bae, Hanyang University,Korea A support vector machine (SVM) provides an optimal separating hyperplane between two classes to be separated. However, the SVM gives only recognition results such as a neural network in a blackbox structure. As an alternative, support vector machine decision tree (SVDT) provides useful information on key attributes while taking a number of advantages of the SVM. We propose an automated parameter selection scheme in SVDT to improve efficiency and accuracy in classification problems. Two practical applications confirm that the proposed methods has a potential in improving generalization and classification error in SVDT.
FP
209
FP-25 PID: 5274 Trend-Weighted Fuzzy Time-series Model for TAIEX Forecasting Ching-Hsue Cheng, National Yunlin University of Science and Technology, Taiwan, R. O. C. Tai-Liang Chen, , National Yunlin University of Science and Technology, Taiwan, R. O. C. Chen-Han Chiang, National Yunlin University of Science and Technology, Taiwan, R. O. C. Time-series models have been used to make reasonably accurate pre-dictions in the areas of weather forecasting, academic enrolment and stock price etc. We propose a methodology which incorporates trend-weighting into the fuzzy time-series models advanced by S.M. Chen and Hui-Kuang Yu. By using actual trading data of Taiwan Stock Index (TAIEX) and the enrolment data of the University of Alabama, we evaluate the accuracy of our trend-weighted, fuzzy, time-series model by comparing our forecasts with those de-rived from Chen’s and Yu’s models. The results indicate that our model surpasses in accuracy those suggested by Chen and Yu. FP-26 PID: 5339 Estimation of Motor Imaginary Using fMRI Experiment Based EEG Sensor Location Sang Han Choi, Kyungpook National University，Korea Minho Lee, Kyungpook National University，Korea Brain computer interface (BCI) is based on brain activity, and controls a computer system through only the imagination or other mental activity. In order to improve the performance of the BCI system based on the scalp EEG, it is critical to get reliable EEG signal, and we need to minimize the noise and artifact from EEG signal. In this paper, we focus on minimizing the artifact besides maximizing the brain activity information related with mental tasks obtained from EEG signals. To make the EEG signal more informative, considering of the mental tasks and the location of EEG sensor is important. Using the fMRI experiment, we found that the supplementary motor area (SMA) is strongly activated whenever the mental tasks are imagination of body movement. Based on this observation, we implement a primitive type of EEG based brain computer interface using the linear discriminant analysis (LDA). FP-27 PID: 5361 Why Not Use an Oracle When You Got One? Ulf Johansson, University of Bor?s, Sweden Tuve L?fstr?m, University of Bor?s, Sweden Rikard K?nig, University of Bor?s, Sweden Lars Niklasson, University of Sk?vde, Sweden The primary goal of predictive modeling is to achieve high accuracy when the model is applied to novel data. For certain problems this requires the use of complex techniques like neural networks or ensembles, resulting in opaque models that are hard or impossible to interpret. For some domains this is unacceptable, since models need to be comprehensible. To achieve comprehensibility, accuracy is often sacrificed by using
FP
210
simpler techniques; a tradeoff termed the accuracy vs. comprehensibility tradeoff. Another, frequently studied, alternative is rule extraction; i.e. the activity where another, transparent, model is generated from the opaque model. In this paper it is argued that existing rule extraction algorithms do not use all information available, and typically should benefit from also using oracle data; i.e. test set instances, together with corresponding predictions from the opaque model. The experiments, using ten publicly available data sets, clearly show that rules extracted using just the oracle data will explain the predictions significantly better than rules extracted in the standard way; i.e. using training data only. FP-28 PID: 5386 Dependency of values of parameters in reinforcement learning for navigation of a mobile robot on the environment Keiji Kamei,Kyushu Institute of Technology，Japan Masumi Ishikawa, Kyushu Institute of Technology, Japan Reinforcement learning has difficulty of the necessity of specifying parameter values in reinforcement learning without prior information. This paper proposes to optimize the parameter values in reinforcement learning by a genetic algorithm with inheritance. In this paper we analyze the dependency of the parameter values on the environment. Computer experiments demonstrate that our proposal succeeded in obtaining optimized parameter values and clarifying the dependency of the values of parameters in reinforcement learning. FP-29 PID: 5424 Comparing Feature Extraction of SVM to Feature Selection of Random MNL and Random Forests for Multiclass Choice Modeling Anita Prinzie, Ghent University, Belgium Dirk Van den Poel, Ghent University, Belgium Choice modeling is a multi-disciplinary application field of data mining. Within marketing multiclass choice modeling, Random Utility (RU) models like MultiNomial Logit (MNL) are well-established due to their microeconomic theoretical underpinnings. Unfortunately, RU models are vulnerable to multicollinearity. As such, they are not well-suited for modeling the choice of instances characterized by many features. To date, RU models lack any feature selection algorithm. We propose a random feature selection procedure integrated within the MNL (Random MNL) classifier as a possible solution. Its predictive performance is compared to that of two machine-learning algorithms, Random Forests (RF) and Support Vector Machines (SVM), known for their ability to handle large feature spaces by random feature selection and feature extraction respectively, but lacking the micro-economic theoretical underpinnings of RU models. The results favor the Random MNL model over both RF and SVM, thereby confirming the appropriateness of RU theory for choice modeling. FP-30 PID: 5448 Application of Neural Networks to Business Bankruptcy Analysis in Thailand Kingkarn Sookhanaphibarn, Chulalongkorn University, Thailand Piruna Polsiri, Dhurakij Pundit University, Thailand Worawat Choensawat, Dhurakij Pundit University, Thailand Frank C. Lin, Chulalongkorn University, Thailand
FP
211
The recent East Asian economic crisis is a lesson one can learn from the absence of effective early warning systems. To serve as a sound early warning signal, the accuracy of a failure prediction model is as important as its robustness over time. This study analyses financial and ownership variables using principal component analysis. It can reduce huge number of financial data of the business bankruptcy prediction problem. Using neural networks for bankruptcy forecasting, the obtained features are fed into neural networks as the input data. Our experiments examine the predictive performance of three neural networks: Learning Vector Quantization, Probabilistic Neural Network, and Feed-forward network with backpropagation learning. All these approached are applied to data sets of 41 Thai financial institutions for the period 1993-2003. FP-31 PID: 5481 Neural Time Series Forecasting and Seasonality: Some New Insights Rohit Dhawan, University of Sydney, Australia Marcus O’Connor, University of Sydney, Australia There has been considerable research in the past on forecasting with artificial neural networks. Such studies involving forecasting with neural networks have either advocated the use of raw data and some have emphasised prior deseasonalisation. While a couple of studies specifically focussing on seasonality in neural time series forecasting affirm that neural networks are inept at modelling the seasonal component of a time series, these undermine the neural network’s ability to approximate any given function, in theory. This paper summarises the debate going on in this area by revisiting existing work and giving new insights on currently used approaches. It will also explore alternative circumstances and techniques that enable neural networks to successfully model seasonal data. FP-32 PID: 5483 Zoomed Clusters Jean-Louis Lassez, Coastal Carolina University, USA Tayfun Karadeniz, Coastal Carolina University, USA Srinivas Mukkamala, Institute for Complex Additive Systems and Analysis, USA We use techniques from Kleinberg’s Hubs and Authorities and kernel functions as in Support Vector Machines to define a new form of clustering. The increase in the degree of non linearity of the kernels leads to an increase in the granularity of the data space and to a natural evolution of clusters into sub-clusters. The algorithm proposed to construct zoomed clusters has been designed to run on very large data sets as found in web directories and bioinformatics. FP-33 PID: 5512 Effect of diffusion weighting and number of sensitizing directions on fiber tracking in DTI Bo Zheng, Nanyang Technological University, Singapore Jagath C. Rajapakse, Nanyang Technological University, Singapore Diffusion Tensor (DT) fiber tracking techniques offer significant potential for studying anatomical connectivity of human brain in vivo. And the reliability and accuracy of fiber tracking results depend on the quality of estimated DT which is determined by parameters of image acquisition protocol. The aim of this paper is to investigate what echo-planar
FP
212
image (EPI) acquisition parameters: the number of sensitizing directions K and diffusion weighting b-value gives the best estimation of DT and shorter scan time. We carried out tracking on synthetic dataset that was artificially corrupted by various levels of Gaussian noise to study the effects of K and b-value on fiber tracking results, and to evaluate the quality of estimated DT. It was found that when K value larger than 13 and b-value larger than 800 smm-2 best estimated DTs. And further increments of K and b-value had no significant effect on quality of estimated DT. FP-34 PID: 5518 Sparse Bump Sonification: A new tool for multichannel EEG diagnosis of mental disorders - application to the detection of the early stage of Alzheimer’s disease. Francois B. Vialatte, BSI RIKEN ABSP Lab, Japan Andrzej Cichocki, BSI RIKEN ABSP Lab, Japan This paper investigates the use of sound and music as a means of representing and analyzing multi-channel EEG recordings. Specific focus is given to applications in early detection and diagnosis of early stage of Alzheimer’s disease. We propose here a novel approach based on multi channel sonification, with a time-frequency representation and sparsification process using bump modeling. The fundamental question explored in this paper is whether clinically valuable information, not available from the conventional graphical EEG representation, might become apparent through an audio representation. Preliminary evaluation of the obtained music score – by sample entropy, number of notes, and synchronous activity – incurs promising results. FP-35 PID: 5541 Clustering Massive High Dimensional Data with Dynamic Feature Maps Rasika Amarasiri, Monash University, Australia Damminda Alahakoon, Monash University, Australia Kate Smith-Miles, Deakin University, Australia This paper presents an algorithm based on the Growing Self Organizing Map (GSOM) called the High Dimensional Growing Self Organizing Map with Randomness (HDGSOMr) that can cluster massive high dimensional data efficiently. The original GSOM algorithm is altered to accommodate for the issues related to massive high dimensional data. These modifications are presented in detail with experimental results of a massive real-world dataset. FP-36 PID: 5579 Exchange Options Pricing with Evolutionary Neural-based Fuzzy Inference Systems Hsing-Wen Wang, National Changhua University of Education, Taiwan Since 1973, Fisher Black and Myron Scholes derived the BSM that provides a short-cut pricing method for options. And then, the option markets and earlier studies take the BSM as the practical model and develop more and more prospering. However, BSM based on many assumptions and constrains such that the price derived with this model was incorrect while compared with the practical prices of the market. Over the past years, many researches in computational intelligence areas reveal that the artificial neural networks (ANNs) pertain excellent learning, high speed computing capabilities, fault-tolerance abilities and the capability of processing non-linear problems to overcome the drawbacks derived from BSM. On the other hand, the adaptive structure neural networks
FP
213
integrated with fuzzy inference systems optimized using Extended-Kalman predictor and modified backpropagation algorithms while considering the fuzzy environments are proposed by Jang (1993). We employ the proposed options pricing model through enhanced neural-fuzzy-based inference systems (ENFIS), whose initial parameters of premise universe can be adjusted systematically by enhanced fuzzy c-means clustering method (EFCM) and programming initially consequence universe with genetic algorithms (GAs) in options pricing and then compared with the BSM. The evidence from empirical studies is using the Deutsche Mark foreign exchange of the Chicago Mercantile Exchange (CME). Results generated by the ENFIS pricing model would be compared with the original BSM using various volatility strategies through pricing error measurement and interpret capability in the research period from 1990 to 1992. The results show that the ENFIS pricing model is superior to the BSM no matter in error degree, variation degree or in interpretation capability.
214
Session Chair Index Last Name First Name Session WANG DeLiang Plenary Speech on Wed KASABOV Nik WA401 AKIMITSU Toshio WA402 KIL Rhee Man WA403 CHO Sungzoon WA404 KANG Min-jae WA405 LUNGA Dalton WA406 IKEDA Kazushi WA407 MURASHIMA Sadayuki WA408 MAK Manwai WB401 LEE Minho WB402 NGUYEN Ha-Nam WB403 VILLMANN Thomas WB404 PANCHAL Rinku WB405 SMALL Michael WB406 LEUNG Frank WB407 MAK Manwai WC401 YAREMCHUK Vanessa WC402 FURUKAWA Tetsuo WC403 BASCLE Benedicte WC404 CHEN Shu-Heng WC405 NAVET Nicolas WC405 FUNG Wai-keung WC406 MINOAHRA Takashi WC407 KING Irwin Plenary Speech on Thr CHAN Laiwan TA401 WANG Rubin TA402 KWOK James TA403 RAJAPAKSE Jagath TA404 HU Xiaolin TA405 JUANG Jih-Gau TA406 KIM Myung Won TA407 ZHANG Kun TA408 USUI Shiro TB401 LEE Minho TB402 ZHANG Kun TB403 BOUZERDOUM Salim TB404 SMALL Michael TB405 ARIK Sabri TB406 ZHOU Yi TB407 MAK Manwai TB408 WANG Jun TC401 ZHANG Qi TC402 SAITO Toshi TC403 YANG Christopher TC404 OH Sanghoun TC405 LIANG Xun TC406 XU Zenglin TC407 HUANG Tingwen TC408
KING Irwin Panel Session on Fri MAK Manwai FA401 FURUKAWA Tetsuo FA402 YANG Shouyuan FA403 HALGAMUGE Saman FA404 ALAHAKOON Damminda FA404 HUANG Kaizhu FA405 MARTIN-MERINO Manuel FA406 KHAN Usman FA407 LEUNG Andrew FA408 CHAN Laiwan FB401 KIMURA Takayuki FB403 YU Eric FB404 SHI Chaojian FB405 LI Chenghua FB406 KUROGI Shuichi FB407 YANG Haixuan FB408
215
Author Index Last Name First Name PID Session Abe Shigeo (5319) FB408 Abe Shigeo (5320) WA404 Abreu Marjory C C (5104) TP Acar Levent (5178) FA403 Acevedo-Mosqueda María Elena (5149) WP Ahn ChangWook (5197) TA405 Ahn Tae-Chon (5043) TP Ahn Tae-Chon (5042) WB407 Aihara Kazuyuki (4840) TA402 Akimitsu Toshio (5284) WA402 Alahakoon Damminda (5541) FP Aliev Rafik (4888) WB406 Aliev Rashad (4888) WB406 Amarasiri Rasika (5541) FP An GaoYun (5013) TC404 Antunes M (4830) TB405 Araujo Carlos A. Paz de (5360) WP Arie Hiroaki (5183) TC408 Arik Sabri (4208) WA408 Assaad Mohammad (5433) WB406 Atiya Amir (5061) TP Atiya Amir F. (5063) WC407 Bae Suk Joo (5250) FP Bai Guo-Qiang (5377) WA408 Bai Hongliang (5540) FA405 Ban Sang-Woo (5331) WB402 Ban Tao (5319) FB408 Ban Tao (5320) WA404 Bao Zheng (4968) WB403 Barman P C (5376) TC407 Bascle B. (5423) WC404 Basu S. K. (4245) TP Basu S. K. (5230) TP Bayarsaikhan Battulga (4828) TA404 Benton Ryan G. (5450) FA407 Bernier O. (5423) WC404 Bertoni Fabiana Cristina (5461) TP Besdok Erkan (4788) FA407 Betechuoh B. Leke (4902) FP Bhatia A. K. (4245) TP Bhatia A. K. (5230) TP Bhaumik Basabi (4277) WA402 Bittencourt Valnaide G. (4793) TB405 Bone Romuald (5433) WB406 Bonner Anthony J. (5179) FB405 Boo Chang-Jin (5202) WP Bouchachia Abdelhamid (5516) TB402 Bouzerdoum A. (4905) TB404 Bouzerdoum Abdesselam (5244) TP Brown Warick (5124) WB406 Bruckner Bernd (5397) FA406 Cai Wei (4986) TC407 Cai Xun (5189) FA407
Cai Xiongcai (4897) WC404 Canuto Anne M P (4917) TB406 Canuto Anne M P (5104) TP Cao Heng (4678) FP Cao Zhiguo (5009) FP Cao Zhitong (4417) WP Cardot Hubert (5433) WB406 Carneiro Raphael V. (5462) FA408 Carvalho Francisco de A.T. de (5406) WB404 Chai Tianyou TC401 Chan Kevin (4658) FP Chan Laiwan FA401 Chan Lai-Wan (5558) TB403 Chan Jonathan H. (5589) TP Chang Bao Rong (5570) WC405 Chang Bao Rong (5378) WP Chang Chuan-Wei (5402) WB406 Chang Chuan-Yu (5359) TP Chang Chuan-Yu (5562) FA405 Chang Hong-Hao (5359) TP Chang Ray-I (4658) FP Chao Ruey-Ming (5543) FA406 Charoenkitkarn Nipon (5589) TP Chau Rowena (5140) FB406 Chen Bo (4968) WB403 Chen Bo-Wei (5502) TA402 Chen Ching-Horng (5502) TA402 Chen Guangju (5036) FP Chen Hung-Ching (Justin) (5452) WA406 Chen Junying (5075) TC405 Chen Kai-Ju (4607) WB404 Chen Luonan (4281) TP Chen Ming (4874) WB403 Chen Peng (4794) TP Chen Rongchang (4881) FP Chen Ruey-Maw (5321) WA405 Chen Sheng-Jyh (4658) FP Chen Shifu (5077) WA407 Chen SongCan (4825) TA404 Chen Tai-Liang (5274) FP Chen Tianping (4817) TC408 Chen Toly (5206) TP Chen Toly (5532) WB407 Chen X.K. (4779) WB402 Chen Xi-Jun (4808) TA406 Chen Xin (4718) WA405 Chen Xucan (5189) FA407 Chen Xue-wen (5135) TA404 Chen Yi-Wei (5402) WB406 Chen Yuehui (4898) FP Chen Yuehui (5121) FP Chen Zhaoqian (5077) WA407 Chen Shi-Huang (5378) WP Chen Shu-Heng (5559) WC405
216
Chen Songcan (4877) TC404 Chen Wanhai (5364) TP Chen Chung-Ming (5504) FB405 Cheng Ching-Hsue (5274) FP Cheng Long (4823) WC406 Cheng Philip E. (5428) WP Chiang Chen-Han (5274) FP Chikagawa Takeshi (4959) TP Chiou Hou-Kai (4896) WC406 Cho Ming-Yuan (4791) FP Cho Sung-Bae (4835) WB405 Cho Sung-Bae (4836) TP Cho Sung-Bae (5440) TC404 Cho Sungzoon (5223) TA403 Cho Sungzoon (5419) WA404 Cho Tae-jun (4643) FP Choensawat Worawat (5448) FP Choi Gyunghyun (5250) FP Choi Jinhyuk (5332) FA408 Choi Jonghwa (5028) WP Choi Jun Rim (5267) WP Choi Sang Han (5339) FP Chow Tommy W. S. (5264) TC407 Chu Chee-hung Henry (5450) FA407 Chu Ming-Hui (5402) WB406 Chuang Cheng-Hung (5428) WP Chuang Cheng-Long (5504) FB405 Chun Myung-Geun (5002) FB404 Chung T.K. (5302) TB408 Cichocki Andrzej (4915) TB403 Cichocki Andrzej (5521) TA408 Cichocki Andrzej (5518) FP Civicioglu P. (4788) FA407 Cong Fengyu (5551) TB403 Cong Fengyu (5551) WP Costa Jose A. F. (4793) TB405 Cottrell M. (5132) WB404 Cox Pedro Henrique (5435) TC403 Cui Gang (4701) TA407 Dawson Michael R. W. (4771) WC402 de Carvalho Aparecido Augusto (5435) TC403 Demirko Askin (5178) FA403 Deng Yanfang (5170) WA403 Dhawan Rohit (5481) FP Dhir Chandra Shekhar (5326) TA408 Dias Stiven S. (5462) FA408 Ding Fan (4418) TP Ding Jundi (4825) TA404 Ding Kang (4816) FB407 Ding Steven X. (4689) WA403 Ding Steven X. (4690) FB408 Dokur Zümray (4159) TP Don An-Jin (4607) WB404 Du Ding (4678) FP Du Ji-Xiang (4301) FB404 Duch Wlodzislaw WA401 Dutra Thiago (4917) TB406
Eastman J. Ronald (4974) WP Eom Jae-Hong (5426) TB405 Er Meng Joo (4814) WB407 Fan Hou-bin (4210) FA406 Fan Shu (4281) TP Fan Zhi-Gang (5173) TB404 Fang Fu-Min (4791) FP Fardin Jr. Dijalma (5462) FA408 Fayed Hatem (5061) TP Fazlollahi Bijan (4888) WB406 Feng Yucai (5307) WP Feraud Raphael (5295) TC407 Fernandez-Redondo Mercedes (4395) TB406 Fernandez-Redondo Mercedes (4396) TB406 Fernandez-Redondo Mercedes (4397) FB408 Filik U. Basaran (4652) WP Fischer T. (5132) WB404 Florian Razvan V. (5411) WP Foresta F. La (5385) FB405 Fu Chaojin (5350) WA408 Fu Shih-Yu (5562) FA405 Fujimoto Katsuhito (5239) FA405 Fujita Kazuhisa (5240) WC402 Fujita Kazuhisa (5287) WC402 Funaya Hiroyuki (4929) WA407 Fung Chun Che (5124) WB406 Furukawa Tetsuo (4457) WC403 Furukawa Tetsuo (5586) WC403 Furukawa Tetsuo (5590) FA402 Furukawa Tetsuo (5587) WC403 Furukawa Tetsuo (5588) WC403 Furukawa Tetsuo (5582) WC403 Fyfe Colin (4969) TC408 Gao Cao (4210) FA406 Gao Cunchen (5389) WA408 Gao Fang (4701) TA407 Gao Fei (5204) TA407 Gao Shao-xia (4210) FA406 Gao Shihai (5120) FA403 Gao Ying (4282) TC405 Garcez Artur S. d’Avila (5462) FA408 Garg Akhil R (4277) WA402 Gedeon Tamas (5124) WB406 Geibel Peter (5097) WP Geweniger T. (5132) WB404 Gong Jianhua (4472) TC406 Gong Jianhua (4545) WA407 Gong Kefei (5006) WP Gong Qin (4608) WP Gu Ren-Min (4239) TP Gu Weikang (5110) TA405 Gu Wenjin (4678) FP Guan Tian (4608) WP Guan Di (5364) TP Guerreiro Ana M. G. (5360) WP Gui Chao (4859) FA403 Guirimov Babek (4888) WB406
217
Guo Aike TC401 Guo Bing (4712) WP Guo G. (4779) WB402 Guo Jun (5004) TA403 Guo Ping (5348) TP Guo Shen-Bo (4301) FB404 Guo Qiang (5364) TP Guo X. C. (5288) TC405 Hagiwara Ichiro (5551) TB403 Hagiwara Ichiro (5551) WP Halgamuge S. (5564) TP Halgamuge Saman K (5538) FA404 Halgamuge Saman K. (5027) WP Halgamuge Saman K. (5537) FA404 Hammer B. (5132) WB404 Hao Zhifeng (4874) WB403 Hashem Sherif (5061) TP Hashemzadeh Mehdi (5186) TP Hashimoto Hideki (4459) TA407 Hashimoto Hideki (4469) TA407 Hattori Motonobu (5349) TB402 He Zhaoshui (5521) TA408 He Pilian (5006) WP Henriques J. (4830) TB405 Herbarth Olf (4588) FA406 Hern´andez-Espinosa Carlos (4395) TB406 Hern´andez-Espinosa Carlos (4396) TB406 Hern´andez-Espinosa Carlos (4397) FB408 Hirayama Daisuke (5287) WC402 Hirooka Seiichi (5240) WC402 Hirose Akira (5284) WA402 Ho Kevin (5346) FA408 Ho Kevin (5346) WP Hong Chuleui (4858) WA405 Hong Jin-Hyuk (4836) TP Hong Jin-Hyuk (4835) WB405 Honma Shun'ichi (4959) TP Horiike Fumiaki (4959) TP Horio Keiichi (5074) FA402 Horio Keiichi (5256) FA402 Horio Keiichi (5583) FA402 Hosaka Ryosuke (4840) TA402 Hoshino Masaharu (5583) FA402 Hosino Tikara (4584) FA408 Hosokawa Kazue (4473) FP Hotta Yoshinobu (5239) FA405 Hou Yuexian (5006) WP Hou Zeng-Guang (4808) TA406 Hou Xiaodi (4971) TC402 Hou Zeng-Guang (4823) WC406 Hsieh Sheng-Ta (5243) TA408 Hsu Arthur (5537) FA404 Hsu Arthur (5538) FA404 Hsu A. L. (5564) TP Hu Changhua (4225) TB406 Hu Wei (5540) FA405 Hu Xiaolin (5025) TA405
Huang Guang-qiu (4960) FP Huang D. (5264) TC407 Huang De-Shuang (4301) FB404 Huang Jau-Chi (4427) FA404 Huang Jau-Chi (4735) TC402 Huang Jie (4875) TP Huang K. (4779) WB402 Huang Kaizhu (5239) FA405 Huang Kou-Yuan (4607) WB404 Huang Mingxiang (4545) WA407 Huang Shian-Chang (5005) WC405 Huang Tingwen (4884) TP Huang Wei (4140) WA406 Huang Yueh-Min (5321) WA405 Huang Zhi-Kai (4301) FB404 Huper Knut (5563) TA408 Hwang Hyung-Soo (5042) WB407 Hwang Kyu-Baek (4780) WA403 Hwang Soochan (4504) FP Hyvarinen Aapo FB401 Ikeda Kazushi (4929) WA407 Ikeguchi Tohru (4840) TA402 Ikeguchi Tohru (5255) FB403 In Myung Ho (5547) TB403 Inoue Shingo (4422) TC402 Inuso G. (5385) FB405 Iqbal Nadeem (5376) TC407 Ishii Kazuo (5590) FA402 Ishikawa Masumi (5386) FP Ishikawa Masumi (5582) WC403 Ishizaki Shun (5349) TB402 Itoh Katsuyoshi (5205) FB407 Izworski Andrzej (5494) WB402 Jang Kyung-Won (5043) TP Jb Arun (5191) WP Jeong Sungmoon (5338) TP Jia Yinshan (5471) TA403 Jian Jigui (4291) TA406 Jiang L. (4163) TA406 Jiang Ning (4690) FB408 Jiang Qi (5531) TP Jiang Weijin (5116) WB406 Jiang Jiayan (4457) WC403 Jianhong Chen (4418) TP Jianjun Li (4418) TP Jin Dongming (5198) WP Jin TaeSeok (4459) TA407 Jin TaeSeok (4469) TA407 Jin Wuyin (4945) WA402 Johansson Ulf (5361) FP Jordaan Jaco (4284) TC406 Ju Fang (5531) TP Juang Jih-Gau (4896) WC406 Juang Jih-Gau (5249) TA406 Jun Sung-Hae (5285) WP Jung Chai-Yeoung (4556) TP Jung Eun Sung (4828) TA404 Jung Soonyoung (5536) FB406 Jung Soonyoung (5536) FA406 Kabe Takahiro (5218) TC403
218
Kambara Takeshi (5287) WC402 Kamei Keiji (5386) FP Kamimura Ryotaro (5137) TC406 Kamimura Ryotaro (4546) FA404 Kamimura Ryotaro (4663) WP Kang Min-Jae (5202) WP Kang Min-jae (5210) WA405 Kang Pilsung (5223) TA403 Kang Rae-Goo (4556) TP Kang Yuan (5402) WB406 Kao Yonggui (5389) WA408 Karabiyik Ali (5565) WP Karadeniz Tayfun (5483) FP Karim M. A. (5564) TP Kasabov Nik TB401 Kasabov Nikola (5286) TA403 Kashimori Yoshiki (5240) WC402 Kashimori Yoshiki (5287) WC402 Katsumata Naoto (4959) TP Kawana Akio (4596) WP Khan Farrukh Aslam (5210) WA405 Khan Usman (5445) FA407 Kil Rhee Man (4583) WA403 Kim Byoung-Hee (4780) WA403 Kim Byung-Joo (4893) FB403 Kim C. H. (4387) TP Kim Choong-Myung (5542) WC402 Kim Dong Hwee (5542) WC402 Kim Eun Ju (5219) TP Kim Eun Ju (5215) FP Kim H. J. (4387) TP Kim Harksoo (4369) WB404 Kim Ho-Chan (5202) WP Kim Ho-chan (5210) WA405 Kim Ho-Joon (5019) TB404 Kim Il Kon (4893) FB403 Kim Kwang-Baek (4637) TB404 Kim Myung Won (5219) TP Kim Myung Won (5261) TA407 Kim Myung Won (5215) FP Kim S. S. (4387) TP Kim Saejoon (4886) TC408 Kim SangJoo (4459) TA407 Kim Sungcheol (4858) WA405 Kim Sungshin (4637) TB404 Kim Tae-Seong (5547) TB403 Kim Yonghwan (4504) FP Kimura Takayuki (5255) FB403 Kinane Andrew (5279) WP Koh Stephen C. L. (4574) WB405 Kong Feng (5227) TB407 Kong Feng (5232) TB407 Konig Rikard (5361) FP Koo Imhoi (4583) WA403 Kouh Jen-Shaing (4658) FP Koyama Ryohei (4423) TB406 Kraipeerapun Pawalai (5124) WB406 Kumar Arun (5470) TA404 Kung Sun-Yuan (5580) TB408 Kung Sun-Yuan (5580) WB401
Kuo Yen-Ting (4427) FA404 Kuo Yen-Ting (5428) WP Kurashige Hiroki (5015) WA402 Kurban M. (4652) WP Kurnaz Mehmet Nadir (4159) TP Kurogi Shuichi (4423) TB406 Kurogi Shuichi (5205) FB407 Kurogi Shuichi (4422) TC402 Kurosaki Osamu (4473) FP Kusiak Magdalena (4842) TP Kuwahara Daisuke (5205) FB407 Kwok James (5487) FB405 Kwon Man-Jun (5002) FB404 Kwon YouAn (5536) FB406 Kyu Phill (4828) TA404 Laaksonen Jorma (5560) FA405 Lai E. M-K. (4597) WA406 Lai E.M-K. (4669) TB402 Lai Hung-Lin (4607) WB404 Lai Kin Keung (4140) WA406 Lai Liang-Bin (4658) FP Lai Weng Kin (4620) WC407 Lai Kin Keung (5520) WC405 Lam James (5524) TB408 Larkin Daniel (5279) WP Lassez Jean-Louis (5483) FP Lee Dae-Jong (5002) FB404 Lee Geehyuk (5332) FA408 Lee Hyo Jong (5001) TP Lee Hyoung-joo (5419) WA404 Lee Hyunjung (4369) WB404 Lee Juho (5019) TB404 Lee Juyeon (5028) WP Lee Kan-Yuan (5243) TA408 Lee Kidong (4903) FP Lee Kwan-Houng (4643) FP Lee Minho (5331) WB402 Lee Minho (5339) FP Lee Sangjoon (5210) WA405 Lee Soo Yeol (5547) TB403 Lee Soo-Young (5326) TA408 Lee Soo-Young TB401 Lee Sungyoung (5547) TB403 Lee Tsair-Fwu (4791) FP Lee Young-Koo (5547) TB403 Lee Yunsik (5047) TC406 Lee Yunsik (5050) WC404 Lee Kin-Hong (5415) WC407 Lee Minho (5338) TP Lee Soo-Young (5376) TC407 Leen Gayle (4969) TC408 Lemaire V. (5423) WC404 Lemaire Vincent (5295) TC407 Leung Chi-sing (5346) FA408 Leung Chi-sing (5346) WP Leung Kwong-Sak (5415) WC407 Li Cheng Hua (4627) FB406 Li Donghai (5120) FA403 Li F. (4163) TA406 Li H. (4779) WB402
219
Li Jianyu (4559) WP Li Jie (5568) FA403 Li Kun-cheng (4609) TB402 Li Li (4892) FB404 Li Ping (4689) WA403 Li Shutao (5487) FB405 Li Sikun (5189) FA407 Li Weihua (4816) FB407 Li Wenhai (5307) WP Li Xiao Ping (4567) WB403 Li Xiaoli (5091) FB405 Li Yangmin (4718) WA405 Li Zhe (4974) WP Li Zhishu (4712) WP Li Ziling (5170) WA403 Li Hua (5348) TP Li Min (4986) TC407 Li Qian-Mu (4981) FB403 Li Wenye (5415) WC407 LI Zhaohui (4282) TC405 Li Zheng (5364) TP Liang Pei-Ji (5041) WA402 Liang Xun (4838) WA406 Liang Xun (4881) FP Liang Y. C. (5288) TC405 Liang Yanxue (5551) TB403 Liang Yanxue (5551) WP Liao Chen (5487) FB405 Liao Guanglan (5291) WP Liao Ling-Zhi (4572) TA404 Liao Ling-Zhi (5046) WB402 Liao Wudai (4305) WP Liao Xiaoxin (4305) WP Lim Chee Peng (4620) WC407 Lim Heuiseok (5328) WP Lim HeuiSeok (5536) FB406 Lim HeuiSeok (5536) FA406 Lim Sungsoo (4836) TP Lim Sungwoo (5445) FA407 Lin Chun-Ling (5243) TA408 Lin Frank (4249) WP Lin Lili (5110) TA405 Lin Lili (5196) WA407 Lin Xiaohong (5116) WB406 Lin Y. (4779) WB402 Lin Yang-Cheng (5574) FB407 Lin Yang-Cheng (5574) TP Lin Frank C. (5448) FP Ling Ping (4478) WB403 Lingras Pawan (4162) TP Liou Cheng-Yuan (4427) FA404 Liou Cheng-Yuan (4735) TC402 Liou Cheng-Yuan (5428) WP Liou Michelle (5428) WP Liu Feng-Yu (4981) FB403 Liu Bingjie (4225) TB406 Liu Chan-Cheng (5243) TA408 Liu Changping (5540) FA405 Liu H. (4163) TA406 Liu Hanxing (4898) FP
Liu Hongwei (4701) TA407 Liu Hongwei (4968) WB403 Liu Hongyan (5227) TB407 Liu Hongyan (5232) TB407 Liu Jing (4323) WA405 Liu Jing (5422) TP Liu Qingshan (5055) TA405 Liu Shiyuan (5291) WP Liu Wei (4875) TP Liu Wen-Kai (5249) TA406 Liu Xinggao (4655) TP Liu Xue (5041) WA402 Liu Yanheng (5078) FP Liu Zhi-Qiang (4691) FA407 Liu Zhi-Yong (5546) FA405 Liu Chunbo (4472) TC406 Liu Huailiang (4282) TC405 Liu Jun (4877) TC404 Lo C.F. (5302) TB408 Lo Shih-Tang (5321) WA405 Lofstrom Tuve (5361) FP Long Chong (5239) FA405 Long Zhi-ying (4609) TB402 López-Yáñez Itzamá (5149) WP Loy Chen Change (4620) WC407 Lu Bao-Liang (5173) TB404 Lu Bin (4881) FP Lu Chiu-Ping (5502) TA402 Lu Huaxiang (4544) FB407 Lu Jie (4609) TB402 Lu Li (5291) WP Lu Wenlian (4817) TC408 Lunga Dalton (4815) WA406 Luo Dijun (4694) WA404 Luo Si-Wei (4572) TA404 Luo Siwei (4559) WP Luo Siwei (5187) FP Luo Si-Wei (5046) WB402 Lursinsap Chidchanok (4249) WP Lv Zehua (5307) WP Lyu Michael R. (5087) TC407 Lyu Michael R. (5412) FA408 Ma Aidong (5187) FP Ma RuNing (4825) TA404 Ma Run-Nian (5377) WA408 Ma Xiaoyan (4544) FB407 Ma Xin (5531) TP Madabhushi Anant (4558) WB405 Maeda Sakashi (5421) TC404 Magdon-Ismail Malik (5452) WA406 Mak Man-Wai (5580) TB408 Mak Manwai WC401 Mak Man-Wai (5580) WB401 Mammone N. (5385) FB405 Man Kim-Fung (4937) TB407 Mandziuk Jacek (4842) TP Mani V. (4387) TP Mao Chengxiong (4281) TP Marques A (4830) TB405
220
Mart´ın-Merino Manuel (5263) FA406 Mart´ın-Merino Manuel (5275) FA404 Marwala T. (4902) FP Marwala Tshilidzi (5220) TP Marwala Tshilidzi (4533) WA406 Marwala Tshilidzi (4815) WA406 Mata Wilson da (4976) FP Meng Fankun (5120) FA403 Meng Ling-Fu (5502) TA402 Meybodi Mohammad. R. (5186) TP Miki Tsutomu (5051) TB408 Mimata Mitsuru (5205) FB407 Min Junki (5440) TC404 Ming Qinghe (5389) WA408 Minohara Takashi (5270) WC407 Mitoma Tetsuro (4794) TP Mizushima Fuminori (5581) FA402 Mogi Ken (5367) TC402 Mogi Ken (5443) TB402 Molter Colin (4496) WA402 Moon Jaekyoung (5338) TP Morabito F.C. (5385) FB405 Morisawa Hidetaka (5287) WC402 Morris Quaid (5179) FB405 Mu˜noz Alberto (5263) FA406 Mukkamala Srinivas (5483) FP Murashima Sadayuki (5373) WA408 Muslim Muhammad Aziz (5582) WC403 Nakagawa Masahiro (4720) TC408 Nakajima Shinichi (4483) WA403 Nakamura Tomohiro (4959) TP Nam KiChun (5536) FB406 Nam KiChun (5536) FA406 Nam Kichun (5542) WC402 Nam Mi Young (4828) TA404 Namikawa Jun (5183) TC408 Naoi Satoshi (5239) FA405 Navet Nicolas (5559) WC405 Neto Adri?ao D.D. (4976) FP Ng G. S. (4574) WB405 Ng S.C. (4558) WB405 Nguyen Ha-Nam (5318) TB405 Nguyen Ha-Nam (5233) WB403 Niklasson Lars (5361) FP Nishi Tetsuo (5004) TA403 Nishida Shuhei (5590) FA402 Nishida Takeshi (4422) TC402 Nishida Takeshi (4423) TB406 Nishida Takeshi (5205) FB407 Nishiyama Yu (4485) FA408 O’Connor Marcus (5481) FP O'Connor Noel (5279) WP Ogata Tetsuya (5183) TC408 Oh Kyung-Whan (5285) WP Oh Sanghoun (5197) TA405 Oh Sung-Kwun (5043) TP Ohashi Fuminori (4959) TP
Ohn Syng-Yup (5318) TB405 Ohn Syng-Yup (5233) WB403 Okabe Yoichi (5284) WA402 Okada Shota (4422) TC402 Okumura Yoshihiro (4473) FP Oliveira Hallysson (5462) FA408 Ölmez Tamer (4159) TP Ong Chong Jin (4567) WB403 Orman Zeynep (4208) WA408 Oshime Tetsunari (4490) FA404 Pan Li (4892) FB404 Panchal Rinku (5200) WB405 Pang Shaoning (5286) TA403 Park Changsu (5328) WP Park Dong-Chul (5047) TC406 Park Dong-Chul (5049) FA407 Park Dong-Chul (5050) WC404 Park Ho-Sung (5043) TP Park Hyung-Min (5326) TA408 Park Junghee (4903) FP Park KiNam (5536) FB406 Park KiNam (5536) FA406 Park Kyeongmo (4858) WA405 Park Soon Choel (4627) FB406 Patel Pretesh B. (4533) WA406 Pei Zheng (5062) TA403 Peng Bo (5009) FP Peng Cheng (4953) TA408 Peng Hong (5062) TA403 Peng Hui (5103) FB408 Peng Lizhi (4898) FP Peng Xiang (5467) TB408 Phung S. L. (4905) TB404 Pi Daoying (4690) FB408 Poel Dirk Van den (5424) FP Polsiri Piruna (5448) FP Prinzie Anita (5424) FP Qian Xiang (4953) TA408 Qiao Hong (5546) FA405 Qin Sheng-Feng (5445) FA407 Qin Zheng (5075) TC405 Quan Zhong-Hua (4301) FB404 Quek C. (4669) TB402 Quek C. (5257) TB407 Quek C. (4574) WB405 Quek C. (4597) WA406 Raicharoen Thanapant (4249) WP Rajapakse Jagath C. (5470) TA404 Rajapakse Jagath C. (5512) FP Rajapakse Jagath C. (5511) WB407 Ramakrishna R.S. (5197) TA405 Ramakrishnan A. G. (5214) TP Rasheed Tahir (5547) TB403 Rattasiri Waratt (5027) WP Ren Guang (5296) WC406 Ribeiro B (4830) TB405 Ribeiro Bernardete (4972) FP Roeder Stefan (4588) FA406 Roh Seok-Beom (5042) WB407 Rolle-Kampczyk
221
Ulrike (4588) FA406 Rom´an Jesus (5275) FA404 Rosen Alan (5425) WP Rosen David B. (5425) WP Ruan QiuQi (5013) TC404 Rui Zhiyuan (4945) WA402 Ryu Joung Woo (5219) TP Ryu Joung Woo (5261) TA407 S¨arel¨a Jaakko (5561) TB403 Saad David (4931) WC407 Saeki Takashi (5051) TB408 Sagheer Alaa (5421) TC404 Saito Toshimichi (5218) TC403 Saito Toshimichi (4490) FA404 Saito Toshimichi (4950) WA407 Sakai Hiroyuki (4473) FP Sakai Ko (4781) WB402 Sakai Yutaka (5015) WA402 Salihoglu Utku (4496) WA402 Samura Toshikazu (5349) TB402 Santana Laura Emmanuella O (5104) TP Saranathan Manojkumar (5214) TP Saraswathi S. (5190) WA404 Sato Naoyuki (4496) WA402 Savran Aydogan (5565) WP scan Zafer (4159) TP Schleif F.-M. (5132) WB404 Sedai Suman (4828) TA404 Senf Alexander (5135) TA404 Seo Jungyun (4369) WB404 Shao Bo (4417) WP Shao Zhuangfeng (4874) WB403 Sharma Satish (4162) TP Shen Hao (5563) TA408 Shen Kai Quan (4567) WB403 Shen Yanjun (4291) TA406 Shen Z.W. (4779) WB402 Shenouda Emad Andrews (5179) FB405 Shi Tielin (4816) FB407 Shi Zhou (4545) WA407 Shi Chaojian (5398) WC404 Shieh Grace S. (5504) FB405 Shin Dongil (5028) WP Shin Dongkyoo (5028) WP Shu Wei (5319) FB408 Shu Zhan (5524) TB408 Silva Catarina (4972) FP Silva Fabio C.D. (5406) WB404 Silva Joyce Q. (5406) WB404 Silva Ivan Nunes da (5461) TP Sinha Neelam (5214) TP Smith A. J. R. (5564) TP Smith-Miles Kate (5140) FB406 Smith-Miles Kate (5541) FP Smola Alexander J. (5563) TA408 Son YoungDae (4469) TA407 Song Tao (4472) TC406 Song Wang-cheol (5210) WA405 Song Zhanjie (4826) FA403
Song Zhihuan (4689) WA403 Sookhanaphibarn Kingkarn (5448) FP Souto Marcilio C. P.de (4793) TB405 Souto Marcilo C P de (4917) TB406 Souza Alberto F. De (5462) FA408 Souza Danilo L. de (4976) FP Souza Renata M.C.R. de (5406) WB404 Sowmya Arcot (4897) WC404 Sugano Shigeki (5183) TC408 Sui Jianghua (5296) WC406 Suk Jung-Hee (5267) WP Sum John (5346) FA408 Sum John (5346) WP Sun Baolin (4859) FA403 Sun Jun (4323) WA405 Sun Jun (5239) FA405 Sun Jun (5422) TP Sun Tsung-Ying (5243) TA408 Sun Youxian (4655) TP Sun Lei (5093) WC404 Sundararajan N. (5190) WA404 Suresh S. (5190) WA404 Suzuki Satoshi (4950) WA407 Tadeusiewicz Ryszard (5494) WB402 Takahashi Norikazu (5004) TA403 Takamatsu Shingo (4483) WA403 Tamukoh Hakaru (5256) FA402 Tan Min (4808) TA406 Tan T. Z. (4574) WB405 Tan Zheng (4986) TC407 Tan Min (4823) WC406 Tan Xiaoyang (4877) TC404 Tanabe Fumiko (5443) TB402 Tanaka Shinya (4423) TB406 Tanaka Takahiro (5074) FA402 Tani Jun (5183) TC408 Taniguchi Rin-Ichiro (5421) TC404 Teddy S. D. (4597) WA406 Teddy S.D. (4669) TB402 Terchi Abdelaziz (5445) FA407 Tettey T. (4902) FP Tettey Thando (5220) TP Thipakorn Bundit (5589) TP Tian Daxin (5078) FP Tian Mei (4572) TA404 Tian Mei (5046) WB402 Tivive Fok Hing Chi (5244) TP Tokunaga Kazuhiro (5587) WC403 Tokunaga Kazuhiro (5588) WC403 Tomisaki Hiroaki (5205) FB407 Tong Hengqing (5204) TA407 Tong Hengqing (5103) FB408 Tong Hengqing (5170) WA403 Tong Xiaofeng (5540) FA405 Torikai Hiroyuki (4490) FA404 Torikai Hiroyuki (5218) TC403 Torres-Sospedra
222
Joaqu´ın (4395) TB406 Torres-Sospedra Joaqu´ın (4396) TB406 Torres-Sospedra Joaqu´ın (4397) FB408 Toyoshima Takashi (5581) FA402 Tran Chung Nguyen (5047) TC406 Tran Chung Nguyen (5050) WC404 Tsai Hsiu Fen (5570) WC405 Tsai Hsiu Fen (5378) WP Tsukada JMinoru (6001) TA402 Tsuruta Nayouki (5421) TC404 Tsuzuki Shinya (4596) WP Tu Kai-Ti (5249) TA406 Tu Yung-Chin (5359) TP Tung W.L. (5257) TB407 Ukil Abhisek (4283) TP Ukil Abhisek (4284) TC406 Usui Shiro (4473) FP Usui Shiro TA401 Vajpai Jayashri (5191) WP Verma Brijesh (5200) WB405 Vialatte Francois B. (5518) FP Villmann Th. (5132) WB404 Wagatsuma Nobuhiko (4781) WB402 Waledzik Karol (4842) TP Wan Huiyan (5181) WB405 Wang Xiao-hai (4960) FP Wang Zhe (4478) WB403 Wang Bin (5398) WC404 Wang Bo (4825) TA404 Wang Chi-Shun (5543) FA406 Wang Cong (4558) WB405 Wang Daojun (4472) TC406 Wang DeLiang WA401 Wang Dianhui (4712) WP Wang Dingwei (4406) TP Wang Dongyun (4305) WP Wang Feng (5120) FA403 Wang Guang-Li (5041) WA402 Wang Haiqing (4689) WA403 Wang Haiqing (4690) FB408 Wang Hongfeng (4406) TP Wang Hsing-Wen (5579) WC405 Wang Hsing-Wen (5579) FP Wang Huaqing (4794) TP Wang J. Z. (4163) TA406 Wang Jian (5078) FP Wang Jie (5181) WB405 Wang Jun (5025) TA405 Wang Jun (5055) TA405 Wang Liang (5187) FP Wang Liansheng (5189) FA407 Wang Peng (5198) WP Wang R. C. (4163) TA406 Wang Rubin (4799) WP Wang Shoujue TA401 Wang Shouyang (4140) WA406 Wang Shouyang (5520) WC405 Wang Tao (5105) TP
Wang Tao (5540) FA405 Wang Xin (4266) WC406 Wang Yan (5121) FP Wang Yea-Ping (5402) WB406 Wang Yi (4691) FA407 Wang Yumei (5471) TA403 Wang Zhongsheng (5350) WA408 Wang C. Y. (5288) TC405 Wang Hsing-Wen (5005) WC405 Wang Tao (5093) WC404 Watanabe Kazuho (4584) FA408 Watanabe Osamu (4596) WP Watanabe Osamu (4657) TC402 Watanabe Sumio (4483) WA403 Watanabe Sumio (4485) FA408 Watanabe Sumio (4584) FA408 Watanapa Saowaluk C. (5589) TP Wei Pengcheng (5159) FP Wei Yaobing (4945) WA402 Wen Wen (4874) WB403 Wickramarachchi Nalin (5027) WP Wilder-Smith Einar P V (4567) WB403 Won Woong-Jae (5331) WB402 Wong K. Y. Michael (4931) WC407 Wong Kok Wai (5124) WB406 Woo Dong-Min (5049) FA407 Woodley Robert S. (5178) FA403 Wright David (5445) FA407 Wu Lu (5389) WA408 Wu M. (4163) TA406 Wu R.H. (4779) WB402 Wu Xia (4609) TB402 Wu Xiu-Ling (5568) FA403 Wu Yan (4239) TP Wu Yunfeng (4558) WB405 Wu C. G. (5288) TC405 XIE Lei (4691) FA407 Xie Shengli (5521) TA408 Xie Shengli (4683) FP Xiong Li (5103) FB408 Xiong Rong (4694) WA404 Xu J X (5554) TA402 Xu Lei (5546) FA405 Xu Wenbo (4323) WA405 Xu Yuetong (4417) WP Xu Yulin (4305) WP Xu Zenglin (5087) TC407 Xu Anbang (5348) TP Xu Man-Wu (4981) FB403 Xu Wenbo (5422) TP Xu Qing (5093) WC404 Xuan Jianping (5291) WP Yamaguchi Nobuhiko (5439) WA404 Yamaguchi Yoko (4496) WA402 Yamaji Kazutsuna (4473) FP Yamakawa Takeshi (5074) FA402 Yamakawa Takeshi (5256) FA402 Yamakawa Takeshi (5583) FA402 Yamazaki Yoshiyuki (6001) TA402
223
Yan Changfeng (4945) WA402 Yan Jianjun (4678) FP Yáñez-Márquez Cornelio (5149) WP Yang Bo (5121) FP Yang Bo (4898) FP Yang Degang (5159) FP Yang Haixuan (5412) FA408 Yang Huaqian (5159) FP Yang Hui (4266) WC406 Yang Hyun-Seung (5019) TB404 Yang Jian (4838) WA406 Yang Jian (4881) FP Yang Jilin (5062) TA403 Yang Shouyuan (4826) FA403 Yang Shuzhong (4559) WP Yang Xiaowei (4874) WB403 Yang Zhirong (5560) FA405 Yang Wen-Chie (4735) TC402 Yao Li (4609) TB402 Yao Li (5044) TA402 Yao Wangshu (5077) WA407 Yao X C (5554) TA402 Yaremchuk Vanessa (4771) WC402 Ye Datian (4953) TA408 Ye Datian (4608) WP Ye Huilin (4894) FP Yeh Chung-Hsing (5574) TP Yeh Chung-Hsing (5140) FB406 Yeh Chung-Hsing (5574) FB407 Yeo Jiyoung (5331) WB402 Yeo Jiyoung (5338) TP Yeung C. H. (4931) WC407 Yeung Sai-Ho (4937) TB407 Yim Hyungwook (5328) WP Yin Kai (5044) TA402 Yin Yuhai (4678) FP Yoon PalJoo (5338) TP Yoshida Fumihiko (5137) TC406 Yoshida Fumihiko (4663) WP You Jiun-De (4607) WB404 Youn Jin-Seon (5267) WP Yu Hui (4291) TA406 Yu Lean (4140) WA406 Yu Shi (4418) TP Yu Xuegang (5078) FP Yu Gisoon (5542) WC402 Yu Lean (5520) WC405 Yuan Haiying (5036) FP Yuan Runzhang (4898) FP Yuan Xu-dong (4210) FA406 Yuan Zhijian (5560) FA405 Yue Z F (5554) TA402 Zdunek Rafal (4915) TB403 Zeng Delu (4683) FP Zhai Yu Zheng (5538) FA404 Zhang Hong (4981) FB403 Zhang Anne (5135) TA404 Zhang Byoung-Tak (5426) TB405 Zhang Byoung-Tak (4780) WA403
Zhang Changshui (5319) FB408 Zhang Daoqiang (4877) TC404 Zhang Haisheng (4838) WA406 Zhang Jiadong (4281) TP Zhang Ke (5373) WA408 Zhang Kun (5558) TB403 Zhang Lihui (4545) WA407 Zhang Li-bao (4862) FP Zhang Liqing (4971) TC402 Zhang Liqing (5568) FA403 Zhang Pu-Ming (5041) WA402 Zhang Qi (5367) TC402 Zhang Wei (5159) FP Zhang Xingzhou (5364) TP Zhang Xun (5198) WP Zhang Y.P. (4779) WB402 Zhang Yan (5105) TP Zhang Yan (5093) WC404 Zhang Yimin (5540) FA405 Zhang YongSheng (5105) TP Zhang Yun-Chu (4808) TA406 Zhang Zhikang (4799) WP Zhang Zhi-Lin (5568) FA403 Zhang Liming (4457) WC403 Zhao Lian-Wei (4572) TA404 Zhao Lian-Wei (5046) WB402 Zhao Qibin (5568) FA403 Zhao Xiao-Jie (5044) TA402 Zhao Yibiao (5187) FP Zhao Yongjun (4875) TP Zheng Hong (4892) FB404 Zheng Hui (4567) WB403 Zheng Hui (4282) TC405 Zheng Bo (5512) FP Zhong Ming (4162) TP Zhong Yixin FA401 Zhong Yixin (4558) WB405 Zhou Chengxiong (5520) WC405 Zhou Chunguang (4478) WB403 Zhou Juan (5214) TP Zhou Juan (5511) WB407 Zhou Li-Zhu (4691) FA407 Zhou Wei (4956) TC403 Zhou Wenhui (5110) TA405 Zhou Xingwei (4826) FA403 Zhou Yi (4814) WB407 Zhou Zhiheng (4683) FP Zhou Weidong (5531) TP Zhu Xiaoyan (5239) FA405 Zhu Jun (4472) TC406 Zhuang Li (5239) FA405 Zou An-Min (4808) TA406