Presentation is loading. Please wait.

Presentation is loading. Please wait.

Bregman Divergences in Clustering and Dimensionality Reduction COMS 6998-4: Learning and Empirical Inference Irina Rish IBM T.J. Watson Research Center.

Similar presentations


Presentation on theme: "Bregman Divergences in Clustering and Dimensionality Reduction COMS 6998-4: Learning and Empirical Inference Irina Rish IBM T.J. Watson Research Center."— Presentation transcript:

1 Bregman Divergences in Clustering and Dimensionality Reduction COMS 6998-4: Learning and Empirical Inference Irina Rish IBM T.J. Watson Research Center Slide credits: Srujana Merugu, Arindam Banerjee, Sameer Agarwal

2 Outline Intro to Bregman Divergences Clustering with Bregman Divergences  k-means: quick overview  From Euclidean distance to Bregman divergences  Some rate-distortion theory Dimensionality Reduction with Bregman Divergences  PCA: quick overview  Probabilistic Interpretation of PCA; exponential family  From Euclidean distance to Bregman divergences Conclusions

3 Distance (distortion) measures in learning Euclidean distance – most commonly used  Nearest neighbor, k-means clustering, least squares regression, PCA, distance metric learning, etc But…is it always an appropriate type of distance? No!  Nominal attributes (e.g. binary)  Distances between distributions Probabilistic interpretation:  Euclidean distance  Gaussian data  Beyond Gaussian? Exponential family distributions  Bregman divergences

4

5

6

7

8 Squared Euclidean distance is a Bregman divergence

9 Relative entropy (i.e., KL-divergence) is another Bregman divergence

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27 Recall Bregman Diverences

28

29

30

31

32 Now, how about generalizing soft clustering Algorithms using Bregman divergences?

33 (natural parameter)

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53 Add a bit of unit-variance Gaussian noise to each point

54 Now remove the original model…

55

56

57

58

59 Remember the exponential family?

60

61

62 Remember Bregman Divergences?

63

64

65

66

67

68

69

70

71

72

73

74

75

76 Discussion

77

78

79


Download ppt "Bregman Divergences in Clustering and Dimensionality Reduction COMS 6998-4: Learning and Empirical Inference Irina Rish IBM T.J. Watson Research Center."

Similar presentations


Ads by Google