1. Machine Learning: Foundations Yishay Mansour Tel-Aviv University

2. Typical Applications Classification/Clustering problems: –Data mining –Information Retrieval –Web search Self-customization –news –mail

3. Typical Applications Control –Robots –Dialog system Complicated software –driving a car –playing a game (backgammon)

4. Why Now? Technology ready: –Algorithms and theory. Information abundant: –Flood of data (online) Computational power –Sophisticated techniques Industry and consumer needs.

5. Example 1: Credit Risk Analysis Typical customer: bank. Database: –Current clients data, including: –basic profile (income, house ownership, delinquent account, etc.) –Basic classification. Goal: predict/decide whether to grant credit.

6. Example 1: Credit Risk Analysis Rules learned from data: IF Other-Delinquent-Accounts > 2 and Number-Delinquent-Billing-Cycles >1 THEN DENAY CREDIT IF Other-Delinquent-Accounts = 0 and Income > $30k THEN GRANT CREDIT

7. Example 2: Clustering news Data: Reuters news / Web data Goal: Basic category classification: –Business, sports, politics, etc. –classify to subcategories (unspecified) Methodology: –consider “typical words” for each category. –Classify using a “distance “ measure.

8. Example 3: Robot control Goal: Control a robot in an unknown environment. Needs both –to explore (new places and action) –to use acquired knowledge to gain benefits. Learning task “control” what is observes!

10. A Glimpse in to the future Today status: –First-generation algorithms: –Neural nets, decision trees, etc. Well-formed data-bases Future: –many more problems: –networking, control, software. –Main advantage is flexibility!

11. Relevant Disciplines Artificial intelligence Statistics Computational learning theory Control theory Information Theory Philosophy Psychology and neurobiology.

12. Type of models Supervised learning –Given access to classified data Unsupervised learning –Given access to data, but no classification Control learning –Selects actions and observes consequences. –Maximizes long-term cumulative return.

13. Learning: Complete Information Probability D 1 over and probability D 2 for Equally likely. Computing the probability of “smiley” given a point (x,y). Use Bayes formula. Let p be the probability.

14. Predictions and Loss Model Boolean Error –Predict a Boolean value. –each error we lose 1 (no error no loss.) –Compare the probability p to 1/2. –Predict deterministically with the higher value. –Optimal prediction (for this loss) Can not recover probabilities!

15. Predictions and Loss Model quadratic loss –Predict a “real number” q for outcome 1. –Loss (q-p) 2 for outcome 1 –Loss ([1-q]-[1-p]) 2 for outcome 0 –Expected loss: (p-q) 2 –Minimized for p=q (Optimal prediction) recovers the probabilities Needs to know p to compute loss!

16. Predictions and Loss Model Logarithmic loss –Predict a “real number” q for outcome 1. –Loss log 1/q for outcome 1 –Loss log 1/(1-q) for outcome 0 –Expected loss: -p log q -(1-p) log (1-q) –Minimized for p=q (Optimal prediction) recovers the probabilities Loss does not depend on p!

17. The basic PAC Model Unknown target function f(x) Distribution D over domain X Goal: find h(x) such that h(x) approx. f(x) Given H find h  H that minimizes Pr D [h(x)  f(x)]

18. Basic PAC Notions S - sample of m examples drawn i.i.d using D True error  (h)= Pr D [h(x)=f(x)] Observed error  ’(h)= 1/m |{ x  S | h(x)  f(x) }| Example (x,f(x)) Basic question: How close is  (h) to  ’(h)

19. Bayesian Theory Prior distribution over H Given a sample S compute a posterior distribution: Maximum Likelihood (ML) Pr[S|h] Maximum A Posteriori (MAP) Pr[h|S] Bayesian Predictor  h(x) Pr[h|S].

20. Nearest Neighbor Methods Classify using near examples. Assume a “ structured space ” and a “ metric ” + + + + - - - - ?

21. Computational Methods How to find an h e H with low observed error. Heuristic algorithm for specific classes. Most cases computational tasks are provably hard.

22. Separating Hyperplane Perceptron: sign(  x i w i ) Find w 1.... w n Limited representation x1x1 xnxn w1w1 wnwn  sign

23. Neural Networks Sigmoidal gates: a=  x i w i and output = 1/(1+ e -a ) Back Propagation x1x1 xnxn

24. Decision Trees x 1 > 5 x 6 > 2 +1+1

25. Decision Trees Limited Representation Efficient Algorithms. Aim: Find a small decision tree with low observed error.

26. Decision Trees PHASE I: Construct the tree greedy, using a local index function. Ginni Index : G(x) = x(1-x), Entropy H(x)... PHASE II: Prune the decision Tree while maintaining low observed error. Good experimental results

27. Complexity versus Generalization hypothesis complexity versus observed error. More complex hypothesis have lower observed error, but might have higher true error.

28. Basic criteria for Model Selection Minimum Description Length  ’(h) + |code length of h| Structural Risk Minimization:  ’(h) + sqrt{ log |H| / m }

29. Genetic Programming A search Method. Local mutation operations Cross-over operations Keeps the “ best ” candidates. Change a node in a tree Replace a subtree by another tree Keep trees with low observed error Example: decision trees

30. General PAC Methodology Minimize the observed error. Search for a small size classifier Hand-tailored search method for specific classes.

31. Weak Learning Small class of predicates H Weak Learning: Assume that for any distribution D, there is some predicate h  H that predicts better than 1/2+ . Weak Learning Strong Learning

32. Boosting Algorithms Functions: Weighted majority of the predicates. Methodology: Change the distribution to target “ hard ” examples. Weight of an example is exponential in the number of incorrect classifications. Extremely good experimental results and efficient algorithms.

33. Support Vector Machine n dimensions m dimensions

34. Support Vector Machine Use a hyperplane in the LARGE space. Choose a hyperplane with a large MARGIN. + + + + - - - Project data to a high dimensional space.

35. Other Models Membership Queries x f(x)

36. Fourier Transform f(x) =  z  z (x)  z (x) = (-1) Many Simple classes are well approximated using large coefficients. Efficient algorithms for finding large coefficients.

37. Reinforcement Learning Control Problems. Changing the parameters changes the behavior. Search for optimal policies.

38. Clustering: Unsupervised learning

39. Unsupervised learning: Clustering

40. Basic Concepts in Probability For a single hypothesis h: –Given an observed error –Bound the true error Markov Inequality Chebyshev Inequality Chernoff Inequality

41. Basic Concepts in Probability Switching from h 1 to h 2 : –Given the observed errors –Predict if h 2 is better. Total error rate Cases where h 1 (x)  h 2 (x) –More refine