exploring the development and exploitation of cutting edge ai technology
TRANSCRIPT
0 copy Copyright 2017 FUJITSU
Fujitsu Forum 2017
FujitsuForum
1 copy Copyright 2017 FUJITSU
Exploring the development and exploitation of cutting-edge AI technology
Shoji Suzuki
Member of the Board
Head of Artificial Intelligence Laboratory
Fujitsu Laboratories Ltd
2 copy Copyright 2017 FUJITSU
Agenda
1 Current status of AI and Fujitsu Lab research strategy
2 Implementation in society of AI technologies
3 Cutting-edge AI technology 「Explainable AI」
3 copy Copyright 2017 FUJITSU
Current status of AI and Fujitsu Lab research strategy
1
4 copy Copyright 2017 FUJITSU
History of Artificial Intelligence
1st AI Boom(SearchInference)
2nd AI Boom(Knowledge)
3rd AI Boom(Machine Learning)
(1956-1960s)
(1980s)
1st AI Winter(lsquo74-rsquo80) 2nd AI Winter(lsquo87-rsquo11)
The birth of AI【lsquo56】 (Dartmouth Conference)
The rush of big AI projects (Expert System)
Games with clear goals can be solved but not practical
Rule Base Lisp
Prolog
(2012-)
Far from the knowledge level of human experts
Deep learning surpassed conventional methods in image recognition【lsquo12】
5 copy Copyright 2017 FUJITSU
Background of the 3rd AI Boom
Evolution of computing Data driven Breakthrough algorithm
The latest GPU 120TFLOPS
Realizes 1000-smartphone computing by one chip
100 GB1 person Sequence data of whole human genome
220 GB1 minute Videos uploaded to Youtube around the world
2 TB1 driving Data generated by various sensors in an automated vehicle
200 TB1 fright Data generated by various sensors in an airplane flight
Data is the new oil of the 21st century
Brilliantly overcame the prejudice that useless
Gains visual and speech recognition functions that can exceed human for the first time with deep learning
2010 2000 1990
Wo
rd e
rro
r ra
te
100
5
10 Using DL
Source Microsoft
Evolution of speech recognition
6 copy Copyright 2017 FUJITSU
Expected AI Related Market
AI related market reaches 598 billion dollars worldwide in 2025 30 for enterprises 70 for consumers
CAGR 50 or more
Ente
rpri
ses
Con
sum
ers
partially edited by Fujitsu Labs
7 copy Copyright 2017 FUJITSU
Solving Practical Social Issues Through AI Current AI applications are still in the element technology level Needs to realize the AI effect and growth through social implementations
Current Applications
Games Speech recognition
translation
Image recognition
Applications required by real world
Mobility System Healthcare System
HR System Social Infrastructure System
Financial System
写真の差し替え by Okamoto
8 copy Copyright 2017 FUJITSU
Strategy to Grow AI Market
① Clearly shows the effect of AI in enterprise market ② Enhance AI market by understanding AIrsquos decision
Shows the effect of AI Understands AIrsquos decision
Collects classifies and organizes events
Understands events and predicts future
Executes by human or AI and feedback its results
Makes decisions and plans actions
Growing Cycle of AI
Visualization
Execution amp Feedback
Analysis amp Prediction
Decision Making
By repeating this cycle AIrsquos performance is demonstrated
AI collaborates smoothly with people and coexists with society
9 copy Copyright 2017 FUJITSU
Implementation in society of AI technologies
2
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
1 copy Copyright 2017 FUJITSU
Exploring the development and exploitation of cutting-edge AI technology
Shoji Suzuki
Member of the Board
Head of Artificial Intelligence Laboratory
Fujitsu Laboratories Ltd
2 copy Copyright 2017 FUJITSU
Agenda
1 Current status of AI and Fujitsu Lab research strategy
2 Implementation in society of AI technologies
3 Cutting-edge AI technology 「Explainable AI」
3 copy Copyright 2017 FUJITSU
Current status of AI and Fujitsu Lab research strategy
1
4 copy Copyright 2017 FUJITSU
History of Artificial Intelligence
1st AI Boom(SearchInference)
2nd AI Boom(Knowledge)
3rd AI Boom(Machine Learning)
(1956-1960s)
(1980s)
1st AI Winter(lsquo74-rsquo80) 2nd AI Winter(lsquo87-rsquo11)
The birth of AI【lsquo56】 (Dartmouth Conference)
The rush of big AI projects (Expert System)
Games with clear goals can be solved but not practical
Rule Base Lisp
Prolog
(2012-)
Far from the knowledge level of human experts
Deep learning surpassed conventional methods in image recognition【lsquo12】
5 copy Copyright 2017 FUJITSU
Background of the 3rd AI Boom
Evolution of computing Data driven Breakthrough algorithm
The latest GPU 120TFLOPS
Realizes 1000-smartphone computing by one chip
100 GB1 person Sequence data of whole human genome
220 GB1 minute Videos uploaded to Youtube around the world
2 TB1 driving Data generated by various sensors in an automated vehicle
200 TB1 fright Data generated by various sensors in an airplane flight
Data is the new oil of the 21st century
Brilliantly overcame the prejudice that useless
Gains visual and speech recognition functions that can exceed human for the first time with deep learning
2010 2000 1990
Wo
rd e
rro
r ra
te
100
5
10 Using DL
Source Microsoft
Evolution of speech recognition
6 copy Copyright 2017 FUJITSU
Expected AI Related Market
AI related market reaches 598 billion dollars worldwide in 2025 30 for enterprises 70 for consumers
CAGR 50 or more
Ente
rpri
ses
Con
sum
ers
partially edited by Fujitsu Labs
7 copy Copyright 2017 FUJITSU
Solving Practical Social Issues Through AI Current AI applications are still in the element technology level Needs to realize the AI effect and growth through social implementations
Current Applications
Games Speech recognition
translation
Image recognition
Applications required by real world
Mobility System Healthcare System
HR System Social Infrastructure System
Financial System
写真の差し替え by Okamoto
8 copy Copyright 2017 FUJITSU
Strategy to Grow AI Market
① Clearly shows the effect of AI in enterprise market ② Enhance AI market by understanding AIrsquos decision
Shows the effect of AI Understands AIrsquos decision
Collects classifies and organizes events
Understands events and predicts future
Executes by human or AI and feedback its results
Makes decisions and plans actions
Growing Cycle of AI
Visualization
Execution amp Feedback
Analysis amp Prediction
Decision Making
By repeating this cycle AIrsquos performance is demonstrated
AI collaborates smoothly with people and coexists with society
9 copy Copyright 2017 FUJITSU
Implementation in society of AI technologies
2
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
2 copy Copyright 2017 FUJITSU
Agenda
1 Current status of AI and Fujitsu Lab research strategy
2 Implementation in society of AI technologies
3 Cutting-edge AI technology 「Explainable AI」
3 copy Copyright 2017 FUJITSU
Current status of AI and Fujitsu Lab research strategy
1
4 copy Copyright 2017 FUJITSU
History of Artificial Intelligence
1st AI Boom(SearchInference)
2nd AI Boom(Knowledge)
3rd AI Boom(Machine Learning)
(1956-1960s)
(1980s)
1st AI Winter(lsquo74-rsquo80) 2nd AI Winter(lsquo87-rsquo11)
The birth of AI【lsquo56】 (Dartmouth Conference)
The rush of big AI projects (Expert System)
Games with clear goals can be solved but not practical
Rule Base Lisp
Prolog
(2012-)
Far from the knowledge level of human experts
Deep learning surpassed conventional methods in image recognition【lsquo12】
5 copy Copyright 2017 FUJITSU
Background of the 3rd AI Boom
Evolution of computing Data driven Breakthrough algorithm
The latest GPU 120TFLOPS
Realizes 1000-smartphone computing by one chip
100 GB1 person Sequence data of whole human genome
220 GB1 minute Videos uploaded to Youtube around the world
2 TB1 driving Data generated by various sensors in an automated vehicle
200 TB1 fright Data generated by various sensors in an airplane flight
Data is the new oil of the 21st century
Brilliantly overcame the prejudice that useless
Gains visual and speech recognition functions that can exceed human for the first time with deep learning
2010 2000 1990
Wo
rd e
rro
r ra
te
100
5
10 Using DL
Source Microsoft
Evolution of speech recognition
6 copy Copyright 2017 FUJITSU
Expected AI Related Market
AI related market reaches 598 billion dollars worldwide in 2025 30 for enterprises 70 for consumers
CAGR 50 or more
Ente
rpri
ses
Con
sum
ers
partially edited by Fujitsu Labs
7 copy Copyright 2017 FUJITSU
Solving Practical Social Issues Through AI Current AI applications are still in the element technology level Needs to realize the AI effect and growth through social implementations
Current Applications
Games Speech recognition
translation
Image recognition
Applications required by real world
Mobility System Healthcare System
HR System Social Infrastructure System
Financial System
写真の差し替え by Okamoto
8 copy Copyright 2017 FUJITSU
Strategy to Grow AI Market
① Clearly shows the effect of AI in enterprise market ② Enhance AI market by understanding AIrsquos decision
Shows the effect of AI Understands AIrsquos decision
Collects classifies and organizes events
Understands events and predicts future
Executes by human or AI and feedback its results
Makes decisions and plans actions
Growing Cycle of AI
Visualization
Execution amp Feedback
Analysis amp Prediction
Decision Making
By repeating this cycle AIrsquos performance is demonstrated
AI collaborates smoothly with people and coexists with society
9 copy Copyright 2017 FUJITSU
Implementation in society of AI technologies
2
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
3 copy Copyright 2017 FUJITSU
Current status of AI and Fujitsu Lab research strategy
1
4 copy Copyright 2017 FUJITSU
History of Artificial Intelligence
1st AI Boom(SearchInference)
2nd AI Boom(Knowledge)
3rd AI Boom(Machine Learning)
(1956-1960s)
(1980s)
1st AI Winter(lsquo74-rsquo80) 2nd AI Winter(lsquo87-rsquo11)
The birth of AI【lsquo56】 (Dartmouth Conference)
The rush of big AI projects (Expert System)
Games with clear goals can be solved but not practical
Rule Base Lisp
Prolog
(2012-)
Far from the knowledge level of human experts
Deep learning surpassed conventional methods in image recognition【lsquo12】
5 copy Copyright 2017 FUJITSU
Background of the 3rd AI Boom
Evolution of computing Data driven Breakthrough algorithm
The latest GPU 120TFLOPS
Realizes 1000-smartphone computing by one chip
100 GB1 person Sequence data of whole human genome
220 GB1 minute Videos uploaded to Youtube around the world
2 TB1 driving Data generated by various sensors in an automated vehicle
200 TB1 fright Data generated by various sensors in an airplane flight
Data is the new oil of the 21st century
Brilliantly overcame the prejudice that useless
Gains visual and speech recognition functions that can exceed human for the first time with deep learning
2010 2000 1990
Wo
rd e
rro
r ra
te
100
5
10 Using DL
Source Microsoft
Evolution of speech recognition
6 copy Copyright 2017 FUJITSU
Expected AI Related Market
AI related market reaches 598 billion dollars worldwide in 2025 30 for enterprises 70 for consumers
CAGR 50 or more
Ente
rpri
ses
Con
sum
ers
partially edited by Fujitsu Labs
7 copy Copyright 2017 FUJITSU
Solving Practical Social Issues Through AI Current AI applications are still in the element technology level Needs to realize the AI effect and growth through social implementations
Current Applications
Games Speech recognition
translation
Image recognition
Applications required by real world
Mobility System Healthcare System
HR System Social Infrastructure System
Financial System
写真の差し替え by Okamoto
8 copy Copyright 2017 FUJITSU
Strategy to Grow AI Market
① Clearly shows the effect of AI in enterprise market ② Enhance AI market by understanding AIrsquos decision
Shows the effect of AI Understands AIrsquos decision
Collects classifies and organizes events
Understands events and predicts future
Executes by human or AI and feedback its results
Makes decisions and plans actions
Growing Cycle of AI
Visualization
Execution amp Feedback
Analysis amp Prediction
Decision Making
By repeating this cycle AIrsquos performance is demonstrated
AI collaborates smoothly with people and coexists with society
9 copy Copyright 2017 FUJITSU
Implementation in society of AI technologies
2
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
4 copy Copyright 2017 FUJITSU
History of Artificial Intelligence
1st AI Boom(SearchInference)
2nd AI Boom(Knowledge)
3rd AI Boom(Machine Learning)
(1956-1960s)
(1980s)
1st AI Winter(lsquo74-rsquo80) 2nd AI Winter(lsquo87-rsquo11)
The birth of AI【lsquo56】 (Dartmouth Conference)
The rush of big AI projects (Expert System)
Games with clear goals can be solved but not practical
Rule Base Lisp
Prolog
(2012-)
Far from the knowledge level of human experts
Deep learning surpassed conventional methods in image recognition【lsquo12】
5 copy Copyright 2017 FUJITSU
Background of the 3rd AI Boom
Evolution of computing Data driven Breakthrough algorithm
The latest GPU 120TFLOPS
Realizes 1000-smartphone computing by one chip
100 GB1 person Sequence data of whole human genome
220 GB1 minute Videos uploaded to Youtube around the world
2 TB1 driving Data generated by various sensors in an automated vehicle
200 TB1 fright Data generated by various sensors in an airplane flight
Data is the new oil of the 21st century
Brilliantly overcame the prejudice that useless
Gains visual and speech recognition functions that can exceed human for the first time with deep learning
2010 2000 1990
Wo
rd e
rro
r ra
te
100
5
10 Using DL
Source Microsoft
Evolution of speech recognition
6 copy Copyright 2017 FUJITSU
Expected AI Related Market
AI related market reaches 598 billion dollars worldwide in 2025 30 for enterprises 70 for consumers
CAGR 50 or more
Ente
rpri
ses
Con
sum
ers
partially edited by Fujitsu Labs
7 copy Copyright 2017 FUJITSU
Solving Practical Social Issues Through AI Current AI applications are still in the element technology level Needs to realize the AI effect and growth through social implementations
Current Applications
Games Speech recognition
translation
Image recognition
Applications required by real world
Mobility System Healthcare System
HR System Social Infrastructure System
Financial System
写真の差し替え by Okamoto
8 copy Copyright 2017 FUJITSU
Strategy to Grow AI Market
① Clearly shows the effect of AI in enterprise market ② Enhance AI market by understanding AIrsquos decision
Shows the effect of AI Understands AIrsquos decision
Collects classifies and organizes events
Understands events and predicts future
Executes by human or AI and feedback its results
Makes decisions and plans actions
Growing Cycle of AI
Visualization
Execution amp Feedback
Analysis amp Prediction
Decision Making
By repeating this cycle AIrsquos performance is demonstrated
AI collaborates smoothly with people and coexists with society
9 copy Copyright 2017 FUJITSU
Implementation in society of AI technologies
2
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
5 copy Copyright 2017 FUJITSU
Background of the 3rd AI Boom
Evolution of computing Data driven Breakthrough algorithm
The latest GPU 120TFLOPS
Realizes 1000-smartphone computing by one chip
100 GB1 person Sequence data of whole human genome
220 GB1 minute Videos uploaded to Youtube around the world
2 TB1 driving Data generated by various sensors in an automated vehicle
200 TB1 fright Data generated by various sensors in an airplane flight
Data is the new oil of the 21st century
Brilliantly overcame the prejudice that useless
Gains visual and speech recognition functions that can exceed human for the first time with deep learning
2010 2000 1990
Wo
rd e
rro
r ra
te
100
5
10 Using DL
Source Microsoft
Evolution of speech recognition
6 copy Copyright 2017 FUJITSU
Expected AI Related Market
AI related market reaches 598 billion dollars worldwide in 2025 30 for enterprises 70 for consumers
CAGR 50 or more
Ente
rpri
ses
Con
sum
ers
partially edited by Fujitsu Labs
7 copy Copyright 2017 FUJITSU
Solving Practical Social Issues Through AI Current AI applications are still in the element technology level Needs to realize the AI effect and growth through social implementations
Current Applications
Games Speech recognition
translation
Image recognition
Applications required by real world
Mobility System Healthcare System
HR System Social Infrastructure System
Financial System
写真の差し替え by Okamoto
8 copy Copyright 2017 FUJITSU
Strategy to Grow AI Market
① Clearly shows the effect of AI in enterprise market ② Enhance AI market by understanding AIrsquos decision
Shows the effect of AI Understands AIrsquos decision
Collects classifies and organizes events
Understands events and predicts future
Executes by human or AI and feedback its results
Makes decisions and plans actions
Growing Cycle of AI
Visualization
Execution amp Feedback
Analysis amp Prediction
Decision Making
By repeating this cycle AIrsquos performance is demonstrated
AI collaborates smoothly with people and coexists with society
9 copy Copyright 2017 FUJITSU
Implementation in society of AI technologies
2
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
6 copy Copyright 2017 FUJITSU
Expected AI Related Market
AI related market reaches 598 billion dollars worldwide in 2025 30 for enterprises 70 for consumers
CAGR 50 or more
Ente
rpri
ses
Con
sum
ers
partially edited by Fujitsu Labs
7 copy Copyright 2017 FUJITSU
Solving Practical Social Issues Through AI Current AI applications are still in the element technology level Needs to realize the AI effect and growth through social implementations
Current Applications
Games Speech recognition
translation
Image recognition
Applications required by real world
Mobility System Healthcare System
HR System Social Infrastructure System
Financial System
写真の差し替え by Okamoto
8 copy Copyright 2017 FUJITSU
Strategy to Grow AI Market
① Clearly shows the effect of AI in enterprise market ② Enhance AI market by understanding AIrsquos decision
Shows the effect of AI Understands AIrsquos decision
Collects classifies and organizes events
Understands events and predicts future
Executes by human or AI and feedback its results
Makes decisions and plans actions
Growing Cycle of AI
Visualization
Execution amp Feedback
Analysis amp Prediction
Decision Making
By repeating this cycle AIrsquos performance is demonstrated
AI collaborates smoothly with people and coexists with society
9 copy Copyright 2017 FUJITSU
Implementation in society of AI technologies
2
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
7 copy Copyright 2017 FUJITSU
Solving Practical Social Issues Through AI Current AI applications are still in the element technology level Needs to realize the AI effect and growth through social implementations
Current Applications
Games Speech recognition
translation
Image recognition
Applications required by real world
Mobility System Healthcare System
HR System Social Infrastructure System
Financial System
写真の差し替え by Okamoto
8 copy Copyright 2017 FUJITSU
Strategy to Grow AI Market
① Clearly shows the effect of AI in enterprise market ② Enhance AI market by understanding AIrsquos decision
Shows the effect of AI Understands AIrsquos decision
Collects classifies and organizes events
Understands events and predicts future
Executes by human or AI and feedback its results
Makes decisions and plans actions
Growing Cycle of AI
Visualization
Execution amp Feedback
Analysis amp Prediction
Decision Making
By repeating this cycle AIrsquos performance is demonstrated
AI collaborates smoothly with people and coexists with society
9 copy Copyright 2017 FUJITSU
Implementation in society of AI technologies
2
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
8 copy Copyright 2017 FUJITSU
Strategy to Grow AI Market
① Clearly shows the effect of AI in enterprise market ② Enhance AI market by understanding AIrsquos decision
Shows the effect of AI Understands AIrsquos decision
Collects classifies and organizes events
Understands events and predicts future
Executes by human or AI and feedback its results
Makes decisions and plans actions
Growing Cycle of AI
Visualization
Execution amp Feedback
Analysis amp Prediction
Decision Making
By repeating this cycle AIrsquos performance is demonstrated
AI collaborates smoothly with people and coexists with society
9 copy Copyright 2017 FUJITSU
Implementation in society of AI technologies
2
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
9 copy Copyright 2017 FUJITSU
Implementation in society of AI technologies
2
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
10 copy Copyright 2017 FUJITSU
Fujitsu AI Technology Brand
Agile amp intense
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
11 copy Copyright 2017 FUJITSU
Fujitsursquos AI Goals
To create AI that collaborates with people and is human centric
To create AI that continuously evolves
To provide AI that can be incorporated in our products and services then deployed
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
12 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
13 copy Copyright 2017 FUJITSU
AI Development Overview
AI platform
AI core technologies
Implementation in society
Applications
Software Acceleration of AI programs as data preprocessing parallel processing distributed processing and so on
Sensing and Recognition
Understanding of users intention emotion environment and situation
Knowledge Processing
Acquisition of knowledge and inference and problem solving
Learning Advancement of machine learning in order to expand application range
AGI Artificial General Intelligence
Service and Embedded Incorporating AI into equipment Development amp verification for applying social infrastructure systems
Governance and Social acceptability
Privacy and security protection of intellectual property rights Social acceptability
Eco-system Ecosystem for business technology and human resource development
Hardware Hardware for high-speed AI computing such as HPC DLU and DAU
Social infra- structure Mobility
Customer center
Logistics
Main- tenance
Manu- facturing
Fintech
Digital marketing
Food
agriculture Health
care
Office living
Explainable AI
Applications to solve social problems in various fields
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
14 copy Copyright 2017 FUJITSU
Extending the Scope of Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction a fundamental issue of machine learning
Time-series data Graph data Image Voice Text
Vehicle Recognition
Company A White male Age 30
Cloth Gray
Backpack Pink
Person Recognition
Handwriting Recognition
Practical application
Fujitsu Laboratoriesrsquo scope Conventional scope
Original (201602)
Leveraged by topological data analysis
Original (201610)
Based on tensor expression
Learning Technology
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
15 copy Copyright 2017 FUJITSU
1 Blade Inspection Image Data
Purpose
Technology
Outcome
Develop ldquoNon-destructive testingrdquo to assist human operators in identifying potential defects
Detection with a combination of Deep Learning Image and Signal processing
Blades can now be checked 75 faster taking just 15 hours
To be put in production in Denmark and UK starting Autumn 2017
demonstration
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
16 copy Copyright 2017 FUJITSU
2 Bridge Inspection
Automatic detection of internal damage for bridges
Topological data analysis (TDA) for time-series data from sensor
Contribution to the optimization of bridge maintenance and management tasks
Vibration data External sensor Converting vibration data into figures
Converting figures into numerical values
degree of abnormality
The degree of abnormality and the degree of change
degree of change TDA
Chaos theory DL
intact damaged
wheel load
concrete slab
external sensor
Purpose
Technology
Outcome
Time-series Data
demonstration
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
17 copy Copyright 2017 FUJITSU
Deep Tensor
FUJITSUrsquos proprietary deep learning technology that can learn graph data representing relationships in the real world
Various relationships
SNS (personal relations)
Compound (element
combination)
Business connections
Existing error back propagation method
Extended error back propagation method
Tensor representation
Graph data
Deep Tensor
Business growth forecast
Fintech
Virtual screening
IT drug development
Malware invasion detection
Cyber attack
hellip
Neural network
Learning both neural network and tensor transformation
Communication relations
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
18 copy Copyright 2017 FUJITSU
3 Security(Malware Attack Detection)
Shorten the detection time several days by experts to a few minutes
Detect the behavior of a malware which is difficult for the expert to find out
the example of graph data that only the new method using Deep Tensor can detect the attack
The number of nodes 223
The number of nodes 249
The number of nodes 343
通信パケット長
起動コマンド種類属性
To shorten the malware attack detection time
Detection of several patterns of attacks by multiple tensor technology Conversion of various types log data to graph data and multi tensor form
DeepTensor
Learn the multiple patterns of attacks
Network log
SNS
Cheical compound
Trading histories
Graph Data
Tensor form +
Tensor form +
Tensor form
Mu
ltple Ten
sor Tech
Deep Learning
Multi Tensor
Con
vert Gra
ph
Da
ta
Graph Data
demonstration
Source IP address
Destination IP address
Source port No
Destination port No
Packet length
Command attribute
Purpose
Technology
Outcome
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
19 copy Copyright 2017 FUJITSU
Modeling Inference Optimization Matching Knowledge Processing and Decision amp Support
Modeling and Inference Modeling and Optimization Matching and Modeling
Heat Stress Level Estimation Ship Fuel Reduction Daycare Center Matching
①Accumulated heat stress status
②Heat stress level decision model
Specialistsrsquo knowledge
Workenvironmental states data Heat
stress level
Machine Learning
Heat stress accumulation status decision algorithm
Ship operational data
Voyage data Engine log climate
Performance model at sea
Learning a statistical model ca
rgo
wei
ght
win
d di
rect
ion
wav
e di
rect
ion
wav
e h
eigh
t
win
d sp
eed
Ship Performance
curr
ent
spee
d curr
ent
dire
ctio
n
agin
g
ship
de
sign
High-dimensional statistical analysis
Accurate
Mathematical Optimization
Prediction
Optimization
Fuel Efficient Routing
Maintenance Timing
Day care centers amp Applicant matching
under applicantsrsquo
requirements
Fair amp efficient
Mathematical model of the relationships of complex requirements
Game theoretical Algorithm
best matching
Risk
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
20 copy Copyright 2017 FUJITSU
4 Heat Stress Level Estimation
Safety management of workers heat stress level
Learning based on the specialists judgment using ambient temperature humidity the pulse rate and the momentum
Realize the safety management of workers under hot weather
The accuracy of heat stress risk level detection is more than 94
Started the business from this summer
The heat stress level is estimated and the alarm is
notified
Pulse rate in working
Pulse rate in a rest
Body heat environmental index
Vital sign sensing band
Specialistsrsquo findings
International standard and index related to heat stress
Heat stress accumulation level dicision algorithm
①Accumulated heat stress
Momentum
①Accumulated heat stress based on work and environmental states ②Presumption of heat stress level based on specialistsrsquo findings
②dicision model
Heat stress level
Risk is high
Risk is middle
Risk is low
Safety
Modeling and Inference
Purpose
Technology
Outcome
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
21 copy Copyright 2017 FUJITSU
5 Ship Fuel Reduction Solve the big issue of ship fuel consumption cost reaches 3 billion dollars per major shipping company
Generate statistical model with every factor learnt from operational data Visualize actual performance at see and select fuel efficient routing for each ship
Fuel consumption reduction is about 5 and estimation error is within 2
Time [sec]
Ship
sp
eed
[kn
ot]
Estimated Measured
Estimating ship performance accurately
Generating statistical model
Collecting operational data
Selecting fuel efficient routing for each ship
cargo weight
wind direction
wave direction
wave height
wind speed
Ship
p
erforman
ce
current speed current
direction
aging
ship design
Hig
h-d
ime
nsio
na
l sta
tistical a
na
lysis
Maintenance Timing Optimization Performance degradation estimation
Fuel efficiency
Optimum maintenance timing
age
VDR=Voyage Data Recorder
demonstration
VDR Data
Engine Log data
Modeling and Optimization
Purpose
Technology
Outcome
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
22 copy Copyright 2017 FUJITSU
Fair and fast matching under applicantsrsquo complex requirements
Mathematical model of the relationships of complex requirements which rapidly calculates the best matching based on game theoretical modeling
Time for admission screening for about 8000 children is reduced from several days to seconds
To be put in production as an optional service for ldquoMICJET MISALIO Child-Rearing Supportrdquo during fiscal 2017
6 Daycare Center Matching Matching and Modeling
Capacity2
Capacity2
Capacity2
1
4
2
5
6
3
Daycare Applicant
Child 1 Priority 1
Child 2 Priority 2
Child 3 Priority 3
Child 4 Priority 4
Meets the Rule
Assignment 1 Daycare A Daycare A Daycare B Daycare B
Assignment 2 Daycare A Daycare B Daycare A Daycare B
Assignment 3 Daycare A Daycare B Daycare B Daycare A
Assignment 4 Daycare B Daycare A Daycare A Daycare B
Assignment 5 Daycare B Daycare A Daycare B Daycare A
Assignment 6 Daycare B Daycare B Daycare A Daycare A
Sibling
Sibling
Purpose
Technology
Outcome
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
23 copy Copyright 2017 FUJITSU
Other Applications
Chatbot for call center
Voice automatic translation Business support
Labor-saving in call centers by 247 service by auto-response
Supporting of communication with foreigners
Automatic classification
AI
Business instruction
Meeting offer
Claim
Just information
Automatic detection of intention of mail sender and summary of contents
Cloud Audio signal
Microphone Speaker
Model enhance
Automatic model adjustment according to noise environment
Edge
Technology for structural semantics extraction by relations of words and high accuracy FAQ search
Learning Voice recognition Translation Voice synthesis
user
Automatic Dialog through messenger
247 Service
Digital Agent for Call Center
Natural Dialog Knowhow
Chatbot FAQ Search
Zinrai Platform
Agents
Supporting office workers for troublesome work such as schedule input
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
24 copy Copyright 2017 FUJITSU
Cutting-edge AI technology 「Explainable AI」
3
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
25 copy Copyright 2017 FUJITSU
Outline
Why Explainable AI ldquoBlack boxrdquo issue must be solved in order to expand social applications of AI
Key technologies to realize Explainable AI Deep Tensor and knowledge graphs
Technological points Explaining ldquoreasonsrdquo and ldquobasisrdquo for conclusions are identified
Specific example Genomic medicine field
Future works
New technology ldquoExplainable AIrdquo
Development of brand new AI that can explain the reasons and basis for AI-generated conclusions
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
26 copy Copyright 2017 FUJITSU
Why Explainable AI
Deep learning is a breakthrough technology for AI To enhance the field which utilize deep learning we solved itrsquos ldquoblack box problemrdquo
Explainable AI 1Rely 2Understand 3Control
Can make new findings
Can fulfill accountability Can improve AI
Human
1 2 3
Empower
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
27 copy Copyright 2017 FUJITSU
Key Technologies to Realize Explainable AI
Deep learning technology that can learn graph data
Deep Tensor
Graph data representing relationships in the real world
Knowledge Graph
LOD
Web
Tensor representation (standardized expression)
Graph data Neural network
Existing error back propagation method
Extended error back propagation method
Malware detection
Cyber attack
Learning both neural network and tensor transformation
Virtual screening
IT drug development
Gene
Mutation
PTPN11
Compound
Gene
Systematic knowledge
(utilized for a role to play)
Onset
Disease
Part of
Protein Expression
Drug class
Target Effect
Disease LEOPARD syndrome
Activation
Part of
Mutation
NM_0028344c174CgtG
Public DB Articles
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
28 copy Copyright 2017 FUJITSU
Technological Points
Inference factor Input Output
Basis formation
Output both findings and reasons (inference factors)
Findings
Knowledge Graph
Inference factor identification
Knowledge graph forms basis from input to findings
1 Explaining the reasons for findings Deep Tensor
2 Explaining the basis for findings
b d a c e f g
Deep Tensor + knowledge graph -gt Reasons and basis of AI judgments
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
29 copy Copyright 2017 FUJITSU
By understanding the cause of the disease largely reduce the time for analysis diagnosis treatment orientation by doctors
Application to Genomic Medicine
A diagnosis (drawing blood)
Gene analysis (next-generation sequencer)
The presentation of the therapeutic drug to personality (genetic
abnormality)
Analyzing and reporting
Ref Hokkaido Univ Hospital(httpwwwhuhphokudaiacjphotnewsdetail00001144html)
Identification of the cause gene medicine investigational search
judgment of the treatment
Web Medical articles Bio DB hellip
Learning from 180000 cases of disease mutation
data
Deep Tensor
Inference of disease
PubMed Medical articles
17million
Bio DB 3million records
b d a c e f
More than 10 billion knowledge built from 17 million medical
articles etc
Causative gene
Gene Ontology
Disease Ontology hellip
Forming medically-proven basis from mutation to
disease
Gene mutation
Gene mutation
demonstration
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
30 copy Copyright 2017 FUJITSU
Example of Basis Paths
Composing a path to be the basis from input ldquomutationrdquo to output ldquodiseaserdquo
Input node (gene mutation)
Output node (disease)
Inference factor of Deep Tensor (partial graph having impact on output)
Node complemented by knowledge graph
Input mutation generates abnormality in gene HGNC795 (ATM) and causes tachycardia (a type of ventricular tachycardia)
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
31 copy Copyright 2017 FUJITSU
Future Works
Aim for collaboration of humans and AI in various fields
Utilization of national projects and industry-academia collaboration
Health care
Co-creation with financial institutions
Finance
Implementation within a company
Corporate
Inferring disease from gene mutation
Providing medically reasoned explanations
Forecasting corporate growth from business performance and
economic indicators
Explaining strengths and weaknesses of companies
Detecting employeesrsquo health change from their activity
data
Explaining factors for health maintenance in work
environment
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
32 copy Copyright 2017 FUJITSU
A safer more prosperous and sustainable world
Human Centric Intelligent Society
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
33 copy Copyright 2017 FUJITSU
Fujitsu Sans Light ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacutethorn
yumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucircuumlyacute
thornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl
Fujitsu Sans Medium ndash abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
0123456789 notrdquopound$^amp()_+-=[]rsquo~ltgt| copyuml~iexclcentcurrenyenbrvbarsectumlordflaquoraquonot-
regmacrdegplusmnsup2sup3microparamiddotcedilsup1ordmfrac14frac12frac34iquestAgraveAacuteAcircAtildeAumlAringCcedilEgraveAEligEacuteEcircEumlIgraveIacuteIcircIumlETHNtildeOgraveOacuteOcircOtildeOumltimesOslashUgraveUacuteUcircUumlYacuteTHORNszligagraveaacuteacircatildeaumlaringaeligccedilegraveeacuteecirceumligraveiacuteicirciumlethntildeograveoacuteocircotildeoumldivideoslashugraveuacuteucirc
uumlyacutethornyumlĐıŒœŠšŸŽžƒʼˆˇˉ˙˚˛˜˝-‒ndashmdash―lsquorsquosbquoldquordquobdquodaggerDaggerbullhellippermillsaquorsaquoolinefrasl⁰⁴⁵⁶⁷⁸⁹₀₁₂₃₄₅₆₇₈₉eurotradeΩrarrpart∆prodsumminusradicinfinintasympnelegesdotlozfifl