artificial intelligence machine learning and deep learning: … · 2019-10-11 · computer vision,...
TRANSCRIPT
Artificial Intelligence Machine learning and Deep
Learning: Trends and Tools
Dr. Shaona Ghosh
@shaonaghosh
What is Machine Learning?
Computer algorithms that learn
patterns in data automatically from
large quantities of data
Use the learnt model for prediction
on new data
Machine Learning: Classification
Source:
Microsoft Azure Machine
Learning
Machine Learning: Regression
Source: Microsoft Azure
Machine Learning
Machine Learning: Housing data
David
Hallac e
t a
l. 2
015
Data drives machine learning?
Data collection infrastructure
sensors, mobile phone, internet connectivity
Data management
Private cloud servers, government clusters
Data pre-processing
Data Annotation
Crowdsourcing, hiring annotators, openly available data
Rise of AI (2012 - present)
•Due to advent of deep (new) neural networks(old) ~ huge models
•Availability of huge amounts of training data ~ millions/ billions of samples
•Powerful computational infrastructure~ 1920 CUDA cores on GPU
Notable Results
Surpassed Human Level Performance in Image Classification or Face Recognition
(Kaiming He et al. 2015, Yaniv Taigman et al, 2014)
Matched Human Level Performance in Machine Translation,Speech Recognition
(Google Translate and Baidu Research)
Beat the world’s top human player in the ancient game of Go
(Google DeepMind)
Adam W. Harley. The AI learns to focus on features
Convolution Neural Networks (CNNs)
Source: Andrej Karpathy Stanford
Network identifies features in image data that human eyes
cannot
Semantic Segmentation
by Shuai Zheng et. al. University of Oxford, ICCV 2015
Instance Segmentation
by Romera Paredes et. al, University of Oxford, 2016
Industry Start-up Sustainability Projects
Ca
pe A
naly
tics (
US
A)
Detect the property condition, square footage, roof condition, solar panels and
other details.
Industry Start-up Sustainability Projects
• Propera (Israel): precision agriculture company that uses
computer vision, deep learning, and agtech to figure out
exactly how much water to deliver to plants in particular
locations so as to improve crop yields while conserving
resources.
• The Climate Corporation (USA) : field health view and
surveillance using deep learning
• Peat (Germany) : precision agriculture, surveillance,
disease diagnostics using deep learning
Will machine learning potentially change
the approaches to development?
Given enough annotated/labelled data is available
• Speed of analysis – terabytes of data in hours
• Accuracy of prediction – human level accuracy or even
better in particular perceptual data: images, text, video,
speech
• Automatically discover features – no more manual and
slow feature engineering, biases
Infrastructure: Hardware
Infrastructure: Hardware
Infrastructure: Distributed Hardware
•Amazon AWS Server with GPU support
•Google Cloud with TPU support
•Google Cloud with GPU support
•Microsoft Azure with ML support
Infrastructure: Software
• Deep Learning Libraries
Infrastructure: Open Data
ImageNet – natural images
Pascal VOC – natural images
OpenStreetMap – geospatial
Biometric Recognition dataset
Uber Ride dataset
DataUSA
EU Gender statistics database
Netherlands National Georegister
Many more
https://deeplearning4j.org/opendata
Infrastructure: Policy
• ASILOMAR AI Principles by Future of Life Institute:
principles to empower people with AI
• Partnership in AI
Amazon, Apple, Google, FB, IBM, Microsoft for best
practices and open platform in AI
• Future of Humanity Institute (Oxford)
Strategic Implications of Openness in AI Development by
Nick Bostrom Report
Leverhulme Centre for the Future of Intelligence (Cambridge)
make the most of machine intelligence
Infrastructure: Policy
• Future of Humanity Institute (Oxford) Report on
When Will AI Exceed Human Performance? Evidence from
AI Experts
Strategic Implications of Openness in AI Development by
Nick Bostrom Technical Report 2016
• Leverhulme Centre for the Future of Intelligence
(Cambridge) brings together the best of human
intelligence so that we can make the most of machine
intelligence
Thanks for listening.