AI & Machine Learning. By Jan Øye Lindroos

Transcription

1 AI & Machine Learning By Jan Øye Lindroos

2 About This Talk Brief introduction to AI: Definition and Characteristics Machine Learning: Types of ML, example algorithms Historical Overview: 1940-Present Present Applications: How AI and ML is used today Future Trends: Where are we in 10 years? Getting Started: How to get started with machine learning

3 Artificial Intelligence: Some Definitions First used (John McCarthy, 1956): "the science and engineering of making intelligent machines." Encyclopedia Britannica: The ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings Tesler s theorem: AI is whatever hasn't been done yet

4 Artificial Intelligence: General Characteristics Perception: Interpret sensory input Reasoning: Deductive/Inductive inference, justify/explain inference Action: Coordination and use of affectors (Robots) Learning: Adapt to changes in environment, improve over time Commonsense: knowledge from human experience Creativity: explore/modify/extend domains Self-awareness: Introspection, consciousness

5 Artificial Intelligence: Types of AI Weak AI: Limited domain Human-level performance on specific tasks Recognizing faces, Playing GO, Understanding speech Strong AI: General purpose AI that can truly reason and becomes self-aware Still sci-fi, but frequently predicted to appear within a decade

6 Artificial Intelligence: Implementation Rule Based Reasoning: Hard coded rules (if/then) Based on expert experience Commercial success in the 80s Machine learning: Learn rules from data Inspired by how humans learn Ex: Artificial Neural Networks

7 Machine Learning Herbert Alexander Simon: Learning is any process by which a system improves performance from experience. Machine Learning is concerned with computer programs that automatically improve their performance through experience. Herbert Simon Turing Award 1975 Nobel Prize in Economics 1978

8 Machine Learning ytrain Supervised Learning f(x) Train Relies on labelled training data DTrain Feature xtrain: Input you always have access to Label ytrain : What you want to predict xtrain x2train Learn a function f(x) y by minimizing Loss f(x)=0.5 ytrain=1 Predict Use f(x) to predict ynew from new data xnew Regression : y is continuous (weight[height]) Classification: y is binary (sale=1, no sale=0) ytrain=0 x1train

9 Machine Learning Over & Underfitting Choice of model complexity Important Underfitting Model used for fit too crude bad accuracy Overfitting Too complex model, fit noise good accuracy, but unlikely to generalise well Test dataset to assess how well model generalizes

10 Machine Learning Neural Networks Perceptron: Artificial neuron Neural Nets: Network of artificial neurons Input layer: Original features x Hidden layers: learned features Output layer: f(x) Train weights w so that the output layer matches labels y GPUs and Big Data allows the use of many layers: Deep Learning

11 Machine Learning Commonly used algorithms Decision Trees/ Random Forest: Automatic if/else separation Choose splits that best separates data Random Forest: Collection of trees k Nearest Neighbours: Classify new data based on closest labeled examples Support Vector Machines: Maximise distance to decision boundary Nonlinearities handled by Kernel trick

12 Machine Learning: Unsupervised Learning Learn structures in unlabeled data Useful for: classification learning features reducing dimensions of input Anomaly detection Examples: Clustering algorithms Principal Component Analysis Autoencoders

13 Machine Learning: Reinforcement Learning Trial/Error based learning: 2. Perform action & evaluate outcome Reward actions with desired outcome 3. Adjust behaviour to maximize reward 1. Similar to how we children learn Applications: Logic games Controlling robot actions

14 Artificial Intelligence: Brief History: Early Years < 1960 Deep Learning 2006 Machine Learning Perceptron (McCulloch & Pitts) First mathematical model of neuron SNARC: Neural network machine (Minsky & Edmonds) Computing Machinery & Intelligence (Alan Turing) Turing test: Requirement for human level AI Learning Machines: Human level AI within 50 years 1990 Renaissance 1980 AI Winter 1970 Middle Ages 1960 First AI workshop in Dartmouth Start of AI as a research field Early years 1940

15 Artificial Intelligence: Brief History: Middle Age Logic Theorist & General Problem Solver (Newell and Simon) Simulation of human problem solving methods Able to prove theorems, Tower of Hanoi puzzle Optimism about strong AI: Machines will be capable, within twenty years, of doing any work that a man can do, Herbert Simon (1960) ELIZA (first chatbot) Rogerian psychotherapist Able to fool some people in Turing test Deep Learning 2006 Machine Learning 1990 Renaissance 1980 AI Winter 1970 Middle Ages 1960 Early years 1940

16 Artificial Intelligence: Brief History: AI Winter Deep Learning 2006 Machine Learning Failed expectations of strong AI Research funding disappeared Perceptron shown to be very limited (Minsky) Research on neural nets came to a halt Progress limited by computer resources Natural Language Machine: 20 words vocabulary Attack by philosophers Chinese room argument (John Searle) Focus on philosophy, rather than technical development 1990 Renaissance 1980 AI Winter 1970 Middle Ages 1960 Early years 1940

17 Artificial Intelligence: Brief History: Renaissance Deep Learning 2006 Machine Learning Practical applications in industry Rise of rule based Expert systems Practical approach to AI research More focus on industrial applications Less focus on mimicking the human mind Robotics influence on AI Strong AI needs a body to sense the world Renewed interest in neural nets and machine learning Complexity and cost of expert systems Computing power & training of NNs (backprop) 1990 Renaissance 1980 AI Winter 1970 Middle Ages 1960 Early years 1940

18 Artificial Intelligence: Brief History: Machine Learning Deep Learning 2006 Machine Learning Shift from knowledge to data driven methods: Expert Systems too costly to maintain The internet: Increased availability of data New ML algorithms invented (SVMs, RFs) State of the art performance Much easier to train than NNs AIs outperforms humans in games: 1992: TD-Gammon, human level backgammon 1997: IBM Deep Blue beats Kasparov at chess 1990 Renaissance 1980 AI Winter 1970 Middle Ages 1960 Early years 1940

19 Artificial Intelligence: Present: Deep Learning, 2006 Pretraining of Multilayer Neural Nets (Hinton et al) Allowed faster training of NN with many layers Led to renewed interest in use of NNs GPUs and Cloud Computing Resources Really deep neural nets computationally feasible Big Data Massive amounts of data from internet use Perfect for deep learning Well suited for typical AI problems Image recognition (CNNs) Natural language processing (RNNs) Able to learns feature from simple raw data Deep Learning 2006 Machine Learning 1990 Renaissance 1980 AI Winter 1970 Middle Ages 1960 Early years 1940

20 Current State: Games Machines consistently beat humans at simple games : DeepMind: GO (4-1 vs. Lee Sedol) DeepMind: Montezuma's Revenge Libratus: Poker Typical methods: Reinforcement Learning Deep neural nets/tree search/game theory Reinforcement learning: Self playing (Go, Poker) Repeated attempts (Video games)

21 Current State: Recommender Systems Customize suggestions to users Suggest related products Use cases: Amazon: Recommend products Netflix: Recommend movies Spotify: Curated playlists Typical methods: Collaborative filtering Find similar content/users k Nearest Neighbors

22 Current State: Image recognition Reach Human level image recognition: Google Brain: Discover cats from watching youtube Facebook: Automatic image labeling for blind Bright Box: Self driving cars trained using video games (GTA V) Typical methods: Deep Convolutional Nets Unsupervised/Supervised/Reinforcement Recurrent Neural Nets for captioning

23 AI at Visma: Ongoing projects In Production: Automatic Ticket Dispatching (SVMs) Automatic routing of s Visma SmartScan Automatic parsing of scanned documents Ongoing: Predicting Sales (Random Forest, knns) Predict whether trial users will buy product Talent Management (Undecided) Smart ranking of job applicants

24 Future Trends: AI100: AI & Life in 2030 Review panel considering likely impact of AI in a typical north american city by the year 2030 Cherry picking: Transportation Health Care Public Safety Workforce Applications for warfare not considered

25 Future Trends: Transportation Self-driving cars commonplace by 2020 Automated Traffic infrastructure: Fewer traffic jams Fewer collisions Car as a service (fewer car owners) Not only cars: drones, boats,...

26 Future Trends: Health AI will support doctors in diagnosis, monitoring and surgery More personalized healthcare using genome and data from wearables Massive increase in health data, allows monitoring, alerting, etc Mobile and communicating robots for in-home elder care

27 Future Trends: Public Safety More widespread surveillance (drones, etc) Use AI to detect anomalies pointing to possible crimes (and police malpractice) Use AI to predict when and where crimes are likely to be committed, and by who Use of AI to prioritizing tasks and allocate resources

28 Future Trends: Workforce short term: AI will likely replace tasks, rather than jobs Will also create new jobs by creating new markets long term: AI gradually invade all employment sectors Require safety nets to protect people against structural shifts in the economy

29 ML: Getting Started Math background: Statistics Linear Algebra Optimization Coding skills: Python, Scala, R, Matlab,... Experiment, Read papers, Go online: Kaggle.com: ML crowdsourcing Arxiv.org: Free article archive Coursera.org: Offers online courses

30 Final tip: Be careful Thanks for your time!

31 Backup

32