Pattern Recognition 2

Transcription

1 Pattern Recognition 2 KNN,, Dr. Terence Sim School of Computing National University of Singapore

2 Outline

3 Outline

4 The Bayes Classifier is theoretically optimum. That is, prob. of error (misclassification), P(error), is smallest, given the set of features. But this requires knowing the true class-conditional pdfs. In reality, we don t know, so we estimate from training data. This may break the optimality, because our training data is not representative of all possible data. Nevertheless, the Bayes Classifier is commonly used. See also: [1]

5 Estimating pdfs One common method is to use histograms. Normalized histogram = pdf.

6 Estimating pdfs Multidimensional histograms are possible. But require a lot of training data to estimate. Heuristic: a D-dimensional histogram needs 10 D training samples. This is the Curse of Dimensionality

7 Naive Bayes To overcome the Curse of Dimensionality, one way is to assume that the D features are statistically independent. Recap: two random variables X, Y are statistically independent iff P(X, Y ) = P(X) P(Y ) Suppose feature vector x = [x 1,..., x D ], then: ω = arg max ω j P(x ω j ) P(ω j ) = arg max ω j P(ω j ) D P(x i ω j ) i=1 In practice, use log probabilities to avoid underflow. This is called the Naive Bayes assumption. Surprisingly, it works well in practice!

8 Face Detection using Naive Bayes [2]

9 Outline

10 K -Nearest Neighbor Instead of using pdfs, why not simply estimate the decision boundary? The K -Nearest Neighbor (KNN) method estimates this implicitly.

11 K -Nearest Neighbor All training data are simply stored, with their class labels. Given a point x to be classified, Select the K nearest neighbors of x. Assign to x the majority label of these K neighbors. Usually, K is odd. Ties can still occur: e.g. 3 classes, and K = 3. This is not optimal classifier, but given enough training data, P(error KNN ) 2 P(error Bayes ) Notion of nearness : Distance metric important. Choice of distance metric affects accuracy. Pros: easy to implement. Cons: cannot scale.

12 Implicit Decision Boundary Each Voronoi cell is a region whose nearest neighbor is the training sample enclosed in the cell.

13 Outline

14 How Good is the Classifier? Test it with another set of labelled data (not training set). ω 1 ω 2 ω 3 ω 4 ω ω ω ω Confusion Matrix The row labels are the true labels of the test data; the columns labels are the ones assigned by the classifier. Diagonal entries are the number of correct classifications. Off-diagonal entries are misclassifications. Ideally, matrix should be diagonal, meaning 0 error. Accuracy = trace/total = 63/96 = 65.6 %

15 Cross Validation To guard against a fluke test. Divide T training samples into N equal bins. Leave out 1 st bin. Train using the rest. Test using 1 st bin. Repeat by leaving out each bin in turn, training using the rest, and testing using the omitted bin. Average the accuracies from all runs. This is called N-fold Cross Validation. Leave-1-out cross validation: when N = T

16 Outline

17 What to Use? Choosing features more of an art than a science. No theory says which feature is best for given problem. Experience, insight, familiarity help. Why not try all possible sets of features? NP-hard! Greedy approach: From F possible features, select 1 that gives best accuracy. From F 1 possible features, select 1 that, when combined with previous feature, gives best accuracy. etc. Common features: edges, lines, color histogram, wavelets, Fourier transform, shape, derivatives.

18 Desirable Properties Easy to compute (efficient) Compact (less storage) Good discriminative power Efficient distance metric Robust to image distortions (e.g. rotation, illumination)

19 Principal Components Analysis () a.k.a. Discrete Karhunen Loeve Transform, Hotelling Transform Let y R k be feature vector computed from image (or another feature vector) x R d, where k d. y = W x (1) W is d k and orthogonal. W to be determined from statistics of x. Let s suppose mean and covariance matrix are: E[x] = 0, E[xx ] = C x We want expected error to be small. How to compute W?

20 Recap Expectation is the mean (average) of random variable x : E[x] = x p(x)dx Variance is the expected squared difference from mean m : Var[x] = E[(x m) 2 ] = E[x 2 ] (E(x)) 2 For vectors, Mean: m = E[x] = [E[x 1 ] E[x 2 ] E[x d ]] Covariance matrix: Var[x] = E[(x m)(x m) ] = E[xx ] mm Note: covariance matrix is symmetric and positive semi-definite

21 Recovered vector x r = Wy Error: ɛ = x x r = x WW x We want small expected error: ɛ 2 = ɛ ɛ = (x WW x) (x WW x) = x x x WW x x WW x + x WW WW x = x x x WW x Note that W W = I.

22 Let k = 1, i.e. W is vector, y is scalar. Then E[ε ε] = E[x x] E[(x w)(w x)] = E[x x] E[w xx w] = E[x x] w E[xx ]w = E[x x] w C x w

23 We need to find w that minimizes E[ɛ ɛ] This is the same as maximizing the 2 nd term on right-hand side: max w w C x w But we should normalize by length of w, so define J = w C x w w w (2) Goal: find w to maximize J

24 Take derivatives and set to 0: dj dw = (w w)2c x w (w C x w)2w (w w) 2 = 0 0 = 2C [ xw w ] w w C x w w 2w w w w Note: term in brackets is J, so we rearrange to get: C x w = J w Eigenvalue problem! Thus, w is eigenvector of C x corresponding to largest eigenvalue (= J).

25 In general, is: y = W (x m) where m = E[x] mean, and W is d k matrix containing the k eigenvectors of Var[x] (covariance matrix) corresponding to the top k eigenvalues. W = w 1 w 2 w k w 1 : First principal component, w 2 : Second principal component, etc.

26 notes is a shift and rotation of the axes. w 1 : direction of greatest elongation w 2 : direction of next greatest elongation, and orthogonal to previous eigenvector; etc. W is orthogonal because C x is symmetric. WW I unless k = d

27 notes Best compression of r.v. x, in the "least-squares error" sense. Guaranteed by the derivation of. De-correlation: y = W (x m) Var[y] = C y = W C x W = Λ diagonal eigenvalue matrix i.e. elements of y = [y 1 y k ] are uncorrelated.

28 notes Dimensionality Reduction: x R d, y R k, k d P d j=1 λ j Typically, choose k so that ratio P k i=1 λ i > 90% "Energy" But how to get C x, m? Statistics of x Given data x 1, x 2,... x N, estimate m, C x Sample mean m = 1 N N i=1 x i Sample covariance matrix: Ĉ x = 1 N (x i m)(x i m) N or Scatter matrix S = or 1 N 1 i=1 N (x i m)(x i m) i=1 N (xi m)(x i m)

29 notes Usually, N d, so Ĉx is rank-deficient, also Ĉx is too large. d d matrix. Computational trick: use inner products. Let A = x 1 m x 2 m x N m, A is d N Then A A is N N. Find eigenvectors/values of A A A Av = λv (AA )Av = λ(av) Av is eigenvector of AA Note that AA = N i=1 (x i m)(x i m) = S, scatter matrix Thus we avoid calculating AA

30 example 1 Perhaps the most famous use of is in face recognition, as exemplified in the classic paper by Turk and Pentland [3]. 2 Start with a set of face images, vectorize them, then compute the. The mean and eigenfaces look like this:

31 example 1 A face image is then a weighted sum of these eigenfaces, + mean. 2 Using more eigenfaces leads to better approximation. Orig. image 1 EF 5 EF 15 EF 25 EF

32 example 1 Each face is thus represented as the weights. 2 Recognition can be done by finding closest weight. 3 What about images of non-faces? Image Approximation

33 Linear Discriminant Analysis () a.k.a. Fisher Linear Discriminant has been used as features in e.g. face recognition. See [3] But is good for pattern representation rather than for pattern discrimination Nevertheless, is still useful for Dimensionality Reduction

34 Linear Discriminant Analysis () will result in w shown in left figure. will result in w shown in right figure.

35 Suppose we have C = 2 classes: ω 1, ω 2 Define the within-class scatter matrix as: S W = C i=1 x ω i (x m i )(x m i ) where m i is the mean of class ω i. Think of this as the sum of the covariance matrix of each class. It measures the spread of each class.

36 Also define the between-class scatter matrix as: S B = (m 1 m 2 )(m 1 m 2 ) This measures the separation of the two classes.

37 Fisher s Criterion is defined as: J F = w S B w w S W w (3) Goal: maximize J F, i.e. find a vector w such that different classes are well separated, while spread in each class is reduced. Compare this with criteria Equation (2) Total scatter matrix S = S W + S B, so is maximizing the spread in the total scatter matrix.

38 It can be shown that the solution to Equation (3) is: S B w = λs W w (4) This is called the Generalized Eigenvalue problem. If S W is invertible, then S 1 W S Bw = λw which is the regular eigenvalue problem involving S 1 W S B

39 Once w is found, feature is computed as: y = w x. Using this feature, we can then classify using, say, KNN. If more than 2 classes, then re-define between-class matrix as: S B = C n i (m i m)(m i m) i=1 where n i is the number of training samples in ω i, and m is the mean of all training samples (ignoring classes).

40 There will be C 1 eigenvectors w i, i = 1... C, each solved using Equation (4). These eigenvectors are not necessarily orthogonal! Because S 1 W B may not be symmetric. These eigenvectors form a linear subspace such that Fisher s Criterion is maximized. Put them into a W matrix, and compute feature as before: y = W x.

41 Outline

42 1 The Bayes Classifier is the theoretical best, but estimating the pdfs is a problem. 2 The K -nearest neighbor classifier is easy to implement, but doesn t scale up well. Also: choice of distance metric important. 3 Deciding what features are best for a given pattern recognition problem is an art, not a science. 4 Two common techniques to compute useful features are: and.

43 References R. Duda, P. Hart, and D. Stork. Pattern Classification, 2nd edition. John Wiley and Sons, H. Schneiderman and T. Kanade. A statistical method for 3d object detection applied to faces and cars. In IEEE Conference on Computer Vision and Pattern Recognition, M. Turk and A. Pentland. Eigenfaces for Recognition. Journal of Cognitive Neuroscience, 3(1), 1991.