CSE 517A MACHINE LEARNING INTRODUCTION Spring 2016 Marion Neumann Contents in these slides may be subject to copyright. Some materials are adopted from Killian Weinberger. Thanks, Killian!
Machine Learning Machine Learning MACHINE LEARNING Traditional CS: Traditional CS: Data Data Program Program Output Output Machine Learning: Machine Learning: Data Data Output Output Computer Computer Computer Computer Program Program 1/20/16 2
Machine Learning MACHINE LEARNING Machine Learning: Traditional CS: Data Output Computer Data Program Computer Output 1/20/16 3
Machine Learning MACHINE LEARNING Training: Testing: Data Output Computer Data Program Computer Output 1/20/16 4
DEFINITION OF ML Formally (Mitchell 1997): A computer program A is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E. Informally: Algorithms that improve on some task with experience. 1/20/16 5
A (VERY BRIEF) HISTORY OF ML A (very brief) History of ML 1/20/16 6
SAMUEL S CHECKER PLAYER (1952) Samuel s Checker Player (1952) Basically Shannon s Minimax Algorithm (traditional AI) Basically Shannon s Minimax Algorithm (traditional AI) Included simple learning algorithm to improve to improve board evaluation. board evaluation. Included simple learning algorithm Player improved over time!! à Player improved over time!! 1/20/16 7
PERCEPTRON (1957, FRANK ROSENBLATT) Provable convergence properties Eventually leads to multilayer perceptron = Neural Networks 1/20/16 8
AI WINTER (1969) Minsky & Papert killed AI Burst huge expectation bubble Funding for AI research collapsed for decades 1/20/16 9
REBIRTH AS MACHINE LEARNING Machine Learning: Originally: Mostly a name game to get funding. Profound difference: ML: Bottom up AI: Top down ML: More practical smaller goals Based on Statistics and Optimization, not Logic 1/20/16 10
TD-GAMMON (1994) Gerry Tesauro (IBM) teaches a neural network to play Backgammon The net plays 100K+ games against itself and beats world champion [Neurocomputation 1994] à Algorithm teaches itself how to play so well!!! 1/20/16 11
DEEP BLUE (1997) IBM s Deep Blue wins against Kasparov in chess. Crucial winning move is made due to Machine Learning (G. Tesauro). 1/20/16 12
IBM WATSON WINS JEOPARDY (2011) 1/20/16 13
POPULAR ML APPROACHES BY TIME 1/20/16 14
ICML 2012 ACCEPTED PAPERS acceptance rate 1/20/16 15
ML SUCCESS STORIES: WEB SEARCH Example: Websearch Example: Websearch 1/20/16 16
ML SUCCESS STORIES: SPAM FILTER 1/20/16 17
ML SUCCESS STORIES: RECOMMENDATIONS 1/20/16 18
ONLINE NEWS PLATTFORMS 1/20/16 19
FRAUD DETECTION 1/20/16 20
VIRTUAL SCREENING Drug design identify structures which are most likely to bind to a drug target ZINC database of purchasable chemical compounds contains approx. 35 million entries à deep learning à graph-based machine learning 1/20/16 21
SOON: AUTONOMOUS CARS Soon: Autonomous Cars 1/20/16 22
When will it stop? WHEN WILL IT STOP? The human brain is one big machine learning machine The human brain is one big learning We know that we can still do a lot better! However, it is hard. Very few people can design new ML algorithms. However, it is hard. Very few people can design new ML algorithms. But But many many people people can use can them! use them! 1/20/16 23
TYPES OF ML Supervised learning: Given labeled examples, find the right prediction of an unlabeled example. (e.g. Given annotated images learn to detect faces.) Unsupervised learning: Given data try to discover similar patterns, structure, low dimensional (e.g. automatically cluster news articles by topic) Reinforcement learning: Try to learn from delayed feedback (e.g. robot learns to walk, fly, play chess) 1/20/16 24
ACTION REQUIRED Join Piazza! Find study group (2-4 students) Project 0 (due 29 th of Jan) Find instructions for Project 0 on course webpage Access your SVN repository Get started on Project 0 à ask TAs if you need help! à project0 will be graded once the autograder is up and running! 1/20/16 25