1. Pattern Classification All materials in these slides were taken from Pattern Classification (2nd ed) by R. O. Duda, P. E. Hart and D. G. Stork, John Wiley & Sons, 2000 with the permission of the authors and the publisher

2. Bayes decision theory febr. 17.

3. 2 Supervised learning: Based on training examples (E), learn a modell which works fine on previously unseen examples. Classification: a supervised learning task of categorisation of entities into predefined set of classes Classification

4. Pattern Classification, Chapter 2 (Part 1) 3

5. 4 P(  j | x) = P(x |  j ). P (  j ) / P(x) Posterior = (Likelihood. Prior) / Evidence Where in case of two categories Posterior, likelihood, evidence

6. Pattern Classification, Chapter 2 (Part 1) 5

7. 6 Decision given the posterior probabilities X is an observation for which: if P(  1 | x) > P(  2 | x) True state of nature =  1 if P(  1 | x) < P(  2 | x) True state of nature =  2 This rule minimizes the probability of the error. Bayesian Decision

8. Pattern Classification, Chapter 2 (Part 1) 7 Bayesian Decision Theory – Generalization Use of more than one feature Use more than two states of nature Allowing actions and not only decide on the state of nature Introduce a loss of function which is more general than the probability of error

9. Pattern Classification, Chapter 2 (Part 1) 8 Let {  1,  2,…,  c } be the set of c states of nature (or “categories”) Let {  1,  2,…,  a } be the set of possible actions Let (  i |  j ) be the loss incurred for taking action  i when the state of nature is  j

10. Bayes decision theory - example Automatic trading (on stock exchanges)  1 : the prices will increase (in the future!)  2 : the prices will be lower  3 : the prices won’t change too much We cannot observe  (latent)!  1: buy  2 : sell x: actual prices (and historical prices) x is observed : how much to lose with an action

11. Pattern Classification, Chapter 2 (Part 1) 10 Overall risk R = Sum of all R(  i | x) for i = 1,…,a Minimizing R Minimizing R(  i | x) for i = 1,…, a for i = 1,…,a Conditional risk

12. Pattern Classification, Chapter 2 (Part 1) 11 Select the action  i for which R(  i | x) is minimum R is minimum and R in this case is called the Bayes risk = best performance that can be achieved!

13. Pattern Classification, Chapter 2 (Part 1) 12 Two-category classification  1 : deciding  1  2 : deciding  2 ij = (  i |  j ) loss incurred for deciding  i when the true state of nature is  j Conditional risk: R(  1 | x) =  11 P(  1 | x) + 12 P(  2 | x) R(  2 | x) =  21 P(  1 | x) + 22 P(  2 | x)

14. Pattern Classification, Chapter 2 (Part 1) 13 Our rule is the following: if R(  1 | x) < R(  2 | x) action  1 : “decide  1 ” is taken This results in the equivalent rule : decide  1 if: ( 21 - 11 ) P(x |  1 ) P(  1 ) > ( 12 - 22 ) P(x |  2 ) P(  2 ) and decide  2 otherwise

15. Pattern Classification, Chapter 2 (Part 1) 14 Likelihood ratio: The preceding rule is equivalent to the following rule: Then take action  1 (decide  1 ) Otherwise take action  2 (decide  2 )

16. Pattern Classification, Chapter 2 (Part 1) 15 Exercise Select the optimal decision where: W = {  1,  2 } P(x |  1 ) N(2, 0.5) (Normal distribution) P(x |  2 ) N(1.5, 0.2) P(  1 ) = 2/3 P(  2 ) = 1/3

17. Pattern Classification, Chapter 2 (Part 2) 16 Therefore, the conditional risk is: “The risk corresponding to this loss function is the average probability error”  Zero-one loss function (Bayes Classifier)

18. Pattern Classification, Chapter 2 (Part 2) 17 Classifiers, Discriminant Functions and Decision Surfaces The multi-category case Set of discriminant functions g i (x), i = 1,…, c The classifier assigns a feature vector x to class  i if: g i (x) > g j (x)  j  i

19. Pattern Classification, Chapter 2 (Part 2) 18 Let g i (x) = - R(  i | x) (max. discriminant corresponds to min. risk!) For the minimum error rate, we take g i (x) = P(  i | x) (max. discrimination corresponds to max. posterior!) g i (x)  P(x |  i ) P(  i ) g i (x) = ln P(x |  i ) + ln P(  i ) (ln: natural logarithm!)

20. Pattern Classification, Chapter 2 (Part 2) 19 Feature space divided into c decision regions if g i (x) > g j (x)  j  i then x is in R i ( R i means assign x to  i ) The two-category case A classifier is a “dichotomizer” that has two discriminant functions g 1 and g 2 Let g(x)  g 1 (x) – g 2 (x) Decide  1 if g(x) > 0 ; Otherwise decide  2

21. Pattern Classification, Chapter 2 (Part 2) 20 The computation of g(x)

22. Discriminant functions of the Bayes Classifier with Normal Density Pattern Classification, Chapter 2 (Part 1) 21

23. Pattern Classification, Chapter 2 (Part 2) 22 The Normal Density Univariate density Density which is analytically tractable Continuous density A lot of processes are asymptotically Gaussian Handwritten characters, speech sounds are ideal or prototype corrupted by random process (central limit theorem) Where:  = mean (or expected value) of x  2 = expected squared deviation or variance

24. Pattern Classification, Chapter 2 (Part 2) 23

25. Pattern Classification, Chapter 2 (Part 2) 24 Multivariate density Multivariate normal density in d dimensions is: where: x = (x 1, x 2, …, x d ) t (t stands for the transpose vector form)  = (  1,  2, …,  d ) t mean vector  = d*d covariance matrix |  | and  -1 are determinant and inverse respectively

26. Pattern Classification, Chapter 2 (Part 3) 25 Discriminant Functions for the Normal Density We saw that the minimum error-rate classification can be achieved by the discriminant function g i (x) = ln P(x |  i ) + ln P(  i ) Case of multivariate normal

27. Pattern Classification, Chapter 2 (Part 3) 26 Case  i =  2. I ( I stands for the identity matrix)

28. Pattern Classification, Chapter 2 (Part 3) 27 A classifier that uses linear discriminant functions is called “a linear machine” The decision surfaces for a linear machine are pieces of hyperplanes defined by: g i (x) = g j (x)

29. Pattern Classification, Chapter 2 (Part 3) 28 The hyperplane is always orthogonal to the line linking the means!

30. Pattern Classification, Chapter 2 (Part 3) 29 The hyperplane separating R i and R j always orthogonal to the line linking the means!

31. Pattern Classification, Chapter 2 (Part 3) 30

32. Pattern Classification, Chapter 2 (Part 3) 31

33. Pattern Classification, Chapter 2 (Part 3) 32 Case  i =  (covariance of all classes are identical but arbitrary!) Hyperplane separating R i and R j (the hyperplane separating R i and R j is generally not orthogonal to the line between the means!)

34. Pattern Classification, Chapter 2 (Part 3) 33

35. Pattern Classification, Chapter 2 (Part 3) 34

36. Pattern Classification, Chapter 2 (Part 3) 35 Case  i = arbitrary The covariance matrices are different for each category (Hyperquadrics which are: hyperplanes, pairs of hyperplanes, hyperspheres, hyperellipsoids, hyperparaboloids, hyperhyperboloids)

37. Pattern Classification, Chapter 2 (Part 3) 36

38. Pattern Classification, Chapter 2 (Part 3) 37

39. Pattern Classification, Chapter 2 38 Exercise Select the optimal decision where: W = {  1,  2 } P(x |  1 ) N(2, 0.5) (Normal distribution) P(x |  2 ) N(1.5, 0.2) P(  1 ) = 2/3 P(  2 ) = 1/3

40. Parameter estimation Pattern Classification, Chapter 3 39

41. Data availability in a Bayesian framework We could design an optimal classifier if we knew: P(  i ) (priors) P(x |  i ) (class-conditional densities) Unfortunately, we rarely have this complete information! Design a classifier from a training sample No problem with prior estimation Samples are often too small for class-conditional estimation (large dimension of feature space!) 1

42. A priori information about the problem E.g. assume normality of P(x |  i ) P(x |  i ) ~ N(  i,  i ) Characterized by 2 parameters Estimation techniques Maximum-Likelihood (ML) and the Bayesian estimations Results are nearly identical, but the approaches are different 1

43. Parameters in ML estimation are fixed but unknown! Best parameters are obtained by maximizing the probability of obtaining the samples observed Bayesian methods view the parameters as random variables having some known distribution In either approach, we use P(  i | x) for our classification rule! 1

44. Use the information provided by the training samples to estimate  = (  1,  2, …,  c ), each  i (i = 1, 2, …, c) is associated with each category Suppose that D contains n samples, x 1, x 2,…, x n ML estimate of  is, by definition the value that maximizes P(D |  ) “It is the value of  that best agrees with the actually observed training sample” 2

45. Optimal estimation Let  = (  1,  2, …,  p ) t and let   be the gradient operator We define l(  ) as the log-likelihood function l(  ) = ln P(D |  ) New problem statement: determine  that maximizes the log-likelihood 2

46. Bayesian Estimation In MLE  was supposed fix In BE  is a random variable The computation of posterior probabilities P(  i | x) lies at the heart of Bayesian classification Goal: compute P(  i | x, D) Given the sample D, Bayes formula can be written

47. Pa tte rn Cl as sifi cat ion, Ch ap ter 1

48. 47 Bayesian Parameter Estimation: Gaussian Case Goal: Estimate  using the a-posteriori density P(  | D) The univariate case: P(  | D)  is the only unknown parameter (  0 and  0 are known!)

49. Reproducing density Identifying (1) and (2) yields:

50. The univariate case P(x | D) P(  | D) computed P(x | D) remains to be computed! It provides: (Desired class-conditional density P(x | D j,  j )) Therefore: P(x | D j,  j ) together with P(  j ) and using Bayes formula, we obtain the Bayesian classification rule:

51. Bayesian Parameter Estimation: General Theory P(x | D) computation can be applied to any situation in which the unknown density can be parametrized: the basic assumptions are: The form of P(x |  ) is assumed known, but the value of  is not known exactly Our knowledge about  is assumed to be contained in a known prior density P(  ) The rest of our knowledge  is contained in a set D of n random variables x 1, x 2, …, x n that follows P(x) 5

52. The basic problem is: “Compute the posterior density P(  | D)” then “Derive P(x | D)” Using Bayes formula, we have: And by independence assumption:

53. MLE vs. Bayes estimation If n→∞ they are equal! MLE Simple and fast (convex optimisation vs. numerical integration) Bayes estimation We can express our uncertainty by P(  ) 52

54. Sumamry Bayes decision theory General framework for probabilistic decision making Bayes classifier Classification is a special decision making (  1 : choose  1 ) Zero-one loss function can be omitted Bayes classifier with zero-one loss with Normal Density

55. Parameter estimation General procedures for densities’ parameter estimation based on a sample (it can be applied beyond Bayes classifier) Bayesian Machine learning: the marrige of Bayesian Decision Theory and Parameter estimation from a training sample Summary