Geek Night 17.0 - Artificial Intelligence and Machine Learning
- 2. Lee Sedol vs. AlphaGo Silver, David, et al. "Mastering the game of Go with deep neural networks and tree search." Nature (2016)
- 3. “At least a decade to go before a computer can beat a human expert” Not very long AGo!
- 4. 1000 = 103 100000000000000000000000000000000000000000000000 = 1047 10000000000000000000000000000000000000000000000000 00000000000000000000000000000000000000000000000000 00000000000000000000000000000000000000000000000000 000000000000000000000 = 10170 Why this assessment?
- 6. AI Winter Series of setbacks from 70s till 90s High Expectations Failure of LISP machines Failure of expert systems
- 7. AI Spring IMAGE: GETTY IMAGES/ISTOCKPHOTO
- 8. AI Spring
- 9. AI Spring Domingos, Pedro. "A few useful things to know about machine learning." Communications of the ACM 55.10 (2012)
- 10. What is ML? “Field of study that gives computers the ability to learn without being explicitly programmed”
- 11. Types of ML problems Machine learning Unsupervised learning Supervised learning Reinforcement learning Regression Classification
- 12. Supervised Learning Spoonfeeding labelled examples Numerical values or Discrete class labels Machine has to be ‘trained’ using a large corpus of ‘training data’
- 13. Regression Training data Hypothesis Choosing optimum ‘hypothesis’ from training data Hypothesis chosen has minimum ‘cost’ Typically used in financial applications, like predicting stock prices or likely monetary value of products
- 14. Classification Height Width Decision boundary Finding decision boundaries based on the labels of the training data Non-linear decision boundaries require complex classifiers like SVMs and neural nets
- 15. Classification applications Spam filtering Optical Character Recognition Pedestrian detection
- 16. Unsupervised Learning (Clustering) Training data is not labelled Grouping based on density (DBSCAN, OPTICS), cluster centers (K-Means) or probability distribution (GMM)
- 17. Clustering applications Grouping similar news items Kharinov, M. "Hierarchical pixel clustering for image segmentation." arXiv preprint (2014). Pixel clustering for segmentation
- 18. Reinforcement Learning Teaching a machine by ‘rewarding’ it for good ‘actions’ and ‘punishing’ it for bad ones Attempt is to explore the entire state space for a problem and get the best actions corresponding to each state, also known as ‘policy’
- 19. Reinforcement Learning applications Reinforcement learning used for AlphaGo
- 20. Deep Learning Capturing abstractions using a multi- level or ‘network’ approach Each level or ‘layer’ composed of many simple processing units The internal abstractions are often the best features to use for the problem, so no feature engineering is required
- 21. Artificial Neural Networks (ANNs) Deep networks composed of artificial neurons Inspired by biological neurons Activation function is typically sigmoid, can be tanh or ReLu The method used to train a network is called ‘backpropagation’ Traditional neural networks with all signals propagating in one direction are called ‘feedforward’ networks Structure of a typical biological neuron Typical artificial neuron
- 22. Artificial Neural Networks (ANNs) contd. Rectifier function Logistic function
- 23. Artificial Neural Networks (ANNs) contd. Sigmoid function Typical feedforward neural network
- 24. Recurrent Neural Networks (RNNs) Hidden layers feed back into themselves Can be used to model sequences and for use as associative memory Can take input sequences of arbitrary length using the concept of ‘attention’
- 25. RNN applications (with links) Automatic music generation (Site has source code link) Handwriting synthesis (Site has paper and source code links) Intelligent personal assistants like Siri, Google Now, Cortana Automatic image captioning Sunspring LSTM that generates poems
- 26. Learning Resources Good courses or tutorials for ML Coursera ML by Andrew Ng Datacamp ML course Udacity Deep Learning Learning by doing Kaggle Topcoder Data science Good video lectures for ML Gilbert Strang lectures on Linear Algebra Nando de Freitas Deep Learning Some people I follow in ML Andrej Karpathy Peter Norvig Alex Graves Fei Fei Li Andrew Ng Some good blogs on ML WildML IAmTrask Karpathy’s blog And finally there’s Google Scholar. Read lots of research papers and try to implement them!
- 27. Thank You Happy Learning :D