Upload
hugo-gaevert
View
467
Download
1
Embed Size (px)
Citation preview
How Will AI Change the Role of the Data Scientist?
Hugo Gävert @hgavert
Helsinki Data Science meet-up 2017-02-16
Who am I?Currently:Chief Data Scientist @ Sanoma
Past: • HUT Infolab • Xtract • Nokia
Hugo Gävert, 2017-02-16
Artificial IntelligenceWorld Goals Use cases Examples
Special purpose AI
Restricted, clear inputs
Well defined, finite
- Recommendation engines,
- Credit scoring, - Insurance claim
handling - Image recognition - Playing games;
chess, go, ping pong, …
- Driving car
- GOFAI, - ML, - ANN / Deep
Learning
- Expert systems - Supervised - Unsupervised - Reinforcement
General AI Open, chaotic, messy inputs
Poorly defined, unconstrained
Requirements: - Reasoning, - communication, - learning new
things - ability to apply
skills to new problems
- Design better AI
- Whole brain simulation?
- Robotic form? - Sensing? - Manipulating the
world?
Super human intelligence?
Hugo Gävert, 2017-02-16
Artificial Super-Intelligence
Human Intelligence
Artificial Intelligence
Inte
lligen
ce /
Perfo
rman
ce
Time
Games Expert tasks Mundane tasks- Checkers, 1952 / 1994 - Backgammon, 1979 - Othello, Chess, 1997 - Jeopardy, 2010 - Go, 2016 - Poker, 2017
- Theorem proving, eq solving - Credit scoring / probability
to default, insurance claim fraud
- Medical diagnosis - Speech to text, translation… - Image recognition
- Natural language / understanding text
- Walking - Object manipulation - Driving cars
Lieutenant Commander Data, year 2338?
Human LevelMachine Intelligence: 10%: 2020 50%: 2040-2050 90%: 2080-2100
Hugo Gävert, 2017-02-16
• Original ideas inspired by brains, but nowadays it’s more engineering for machine learning tasks.
• Artificial Neural Network ≈ Layers of connected simple neurons • Multiple different architectures for different uses
Neural Networks?
A cartoon drawing of a biological neuron (left) and its mathematical model (right). Stanford CS231n: Convolutional Neural Networks for Visual Recognition
Hugo Gävert, 2017-02-16
Why Deep Learning?• Rebranded artificial neural networks, so what is different now?
Big Data - Text, images, video - Large annotated data
sources, like images155k words, 117k senses
14M images, 1M BBoxes, 22k synsets
Computational power
Some new algorithms; ReLU, dropouts, initializations, ConvNets
-4 -3 -2 -1 0 1 2 3 4-1
1
-4 -3 -2 -1 0 1 2 3 4-1
1-4 -3 -2 -1 0 1 2 3 4
-1
1
Hugo Gävert, 2017-02-16
Deep Belief Networks• 2006, Geoff Hinton: A Fast Learning Algorithm for Deep Belief Networks
• First major results in 2009 in Acoustic Model using Deep Belief Networks —> Speech recognition
• What is it? • Multilayer feedforward network with • Input layer • Many hidden layers • Output layer • Training…
Train as RBM
Train as RBMTrain with backpropagation
Hugo Gävert, 2017-02-16
From feature engineering to feature learning
Input OutputHand
designed program
Rule-based AI
Trained classifierInput Output
Hand designed features
Classic ML
Features Trained classifierInput OutputRepresentation
Learning
Simple features
Mid level abstract features
Trained classifierInput Output
High level abstract features
DeepLearning
Hugo Gävert, 2017-02-16
• Deep Belief Networks have largely been replaced by convolutional networks for image recognition • Architecture, layers:
• Input (width, height, depth = RGB) • Convolutional layer
• Neuron calculates convolution of the weights over the local image area • N filters with size (width, height, N) • Relu activation layer
• Pooling layer• Downsampling along the spatial width and height dimension
• Fully connected layer (output: 1 x 1 x num of classes) • The conv + relu + pooling layers are repeated.
• Of course, other architectures also…
Convolutional networks?
Hugo Gävert, 2017-02-16
Convolutional networks - What is deep?• AlexNet, 2012
• ImageNet challenge, top 5 error rate 16% (previous 26%) • 5 conv, max-pooling, drop-out layers, 3 fully connected
• ZF Net, 2013 • Top 5 error rate 11.2% • Similar architecture, only 10% of training data • DeConvNet - visualisation of the layers
• VGG Net, 2014 • Top 5 error rate 7.3% • 19 layers, but simple 3x3 convolution and 2x2 max pooling • CNNs need to be deep, but otherwise simple
• GoogLeNet, 2015 • Top 5 error rate 6.7% • 22 layers, but has inception-modules that do work in parallel
• Microsoft ResNet, 2015 • Top 5 error rate 3.6% (better than human) • 152 layers, ultra deep
Hugo Gävert, 2017-02-16
Speech Recognition at Google
Brandon Ballinger: Deep Learning and the Dream of AI, Strata 2013Jaitly et al (2012), Application of pretrained deep neural networks to LVSRHugo Gävert, 2017-02-16
Chatbots and AI• Speech recognition ok
• Natural language understanding needs work
• Logic • If … then… • No memory in session
• Behavior / approach • Reactive, just answers
questions • Proactive would be helpful…
Hugo Gävert, 2017-02-16
Products you should test / use• Google APIs
• Machine learning platform (Deep Learning: TensorFlow)
• Natural Language API • Speech API • Translation API • Vision API
• IBM Watson analytics…
• Also, some of the famous image ConvNets are downloadable in pre-trained format
• MS Azure ML (Cortana analytics, cognitive services) • Deep Learning: CNTK • Vision: Face API, Emotion API,
Computer Vision API, Content Moderation API
• Recommendations API, Academic knowledge API, Entity linking API, Anomaly Detection
• Language: Text Analysis, Web Language Model, spell checking, translation
• Speech: Speech to text, speaker identification, translation
Hugo Gävert, 2017-02-16
So is AI going to take the job of Data Scientists?• Yes, absolutely• Why?
• We, the data scientists, are building the AI - we’re lazy, we build AI to do our job…
• Harder to build the robots (or cars, trucks, flying machines) than to just run the AI inside computer. The early use cases will be confined in the computers.
• When? • Not very soon…
Hugo Gävert, 2017-02-16
What does typical data science project look like?
Business understanding
Data understanding and quality
Data pre-processing
Feature engineering
Modeling
Evaluation
Production deployment
Hugo Gävert, 2017-02-16
What does typical data science project look like?
Business understanding
Data understanding and quality
Data pre-processing
Feature engineering
Modeling
Evaluation
Production deployment
Data collection design
Monitoring, control
Feature learning
Deep Learning architecture
Communications, internal consulting
How do we get representative data for
the network?
Ok, images easy - how about others?
Does it work?
Still expected results?
Fraudulent use?
What is this Black Box?APIs
Hugo Gävert, 2017-02-16
Recommendations for Data Scientists• Keep on doing what you do • Evolve with the world • You still need
• Math; stats, probabilities, linear algebra… • Algorithms and data structures
• You also need now • Deep Learning (hype!) • More communications skills • Software writing & engineering skills (APIs)
• Google and Stack Overflow helps…
Hugo Gävert, 2017-02-16
Recommendations for companies• Data
• Create data strategy; collect, store and make data available • Data is key business asset in building AI capability. Deep
Learning needs data in training. Software can be replicated, but data cannot - if a business has data, then it’s already in better position than competitors.
• Hire talent• AI models need to be customized for the business need,
application and context. • Downloading open source software is not enough.
Applying it is far from trivial. The APIs solve only specific problems and are too much black boxes.
• You need to be able to explain the models to customers - specially in the legal, finance, insurance, health etc. business.
“The best ideas come from the guys closest to the data.”
Todd Holloway Head of Data Science at Trulia.
Hugo Gävert, 2017-02-16