1. Hierarchical Subquery Evaluation for Active Learning on a Graph Oisin Mac Aodha, Neill Campbell, Jan Kautz, Gabriel Brostow CVPR 2014 University College London 1

2. Cat Dog Horse 2 Large Image Collections https://www.flickr.com/photos/cmichel67

3. Large Image Collections https://www.flickr.com/photos/cmichel67 Cat Dog Horse Labeling large image collections is tedious 3

4. Acquiring Annotations 4 https://www.flickr.com/photos/usnavyhttps://www.flickr.com/photos/rdecom Crowdsourcing Specialized Knowledge Expert time is valuable!

5. 5 Active Learning Oracle AL Algorithm User Query Label Unlabeled Dataset

6. Number of user queries Test Accuracy 1 0 6 Learning Curves

7. Number of user queries 1 0 7 Learning Curves Test Accuracy

8. Number of user queries 1 0 8 Learning Curves Test Accuracy

9. Number of user queries 1 0 9 Learning Curves Test Accuracy

10. Learning Curves Number of user queries 1 0 10 Test Accuracy

11. Learning Curves Number of user queries 1 0 We want the largest area under the learning curve 11 Test Accuracy

12. Learning Curves 1 0 12 Test Accuracy The number of unlabeled images can be very large!

13. 13 Active Learning Wish List

14. Fast updating of classifier for interactive labeling 14 Active Learning Wish List

15. Fast updating of classifier for interactive labeling Exploit structure in unlabeled data 15 Active Learning Wish List

16. Fast updating of classifier for interactive labeling Exploit structure in unlabeled data Consistent performance across different datasets 16 Active Learning Wish List

17. Fast updating of classifier for interactive labeling Exploit structure in unlabeled data Consistent performance across different datasets Make the most of the expert’s time 17 Active Learning Wish List Graph Based Semi-Supervised Learning Perplexity Graph Construction Our Hierarchical Subquery Evaluation

18. 18 Related Work Video Segmentation Fathi et al. BMVC 2011 Action Detection Bandla and Grauman ICCV 2013 Gaussian Random Fields Zhu et al. ICML 2003 Semantic Segmentation Vezhnevets et al. CVPR 2012 RALF: Reinforced Active Learning Ebert et al. CVPR 2012 … Image Classification Kapoor et al. ICCV 2007 …

19. xi xi φ(φ( ) = 19 Supervised Classification

20. xj xj φ(φ( ) = 20 Supervised Classification

21. 21 Supervised Classification

22. 22 Supervised Classification Decision Boundary

23. Semi-supervised learning using Gaussian fields and harmonic functions X. Zhu, Z. Ghahramani, J. Lafferty ICML 2003 F i = P(f(x i ) == class1) 23 w ij Semi-Supervised Learning

24. 24 F i = P(f(x i ) == class1) w ij

25. Graph Construction 25 Stochastic neighbor embedding G. Hinton and S. Roweis NIPS 2002

26. 26 Graph Active Learning

27. Example 2 Class Graph 27

28. Example 2 Class Graph 28 Ground Truth

29. Example 2 Class Graph 29 Active Learning Strategies

30. Random 30

31. Active Learning Strategies Random Exploration – clusters 31

32. Active Learning Strategies Random Exploration – clusters Exploitation – uncertainty 32

33. Active Learning Strategies Random Exploration – clusters Exploitation – uncertainty 33

34. Active Learning Strategies Random Exploration – clusters Exploitation – uncertainty RALF – explore or exploit 34 Ralf: A reinforced active learning formulation for object class recognition S. Ebert, M. Fritz, and B. Schiele CVPR 2012

35. Active Learning Strategies Random Exploration – clusters Exploitation – uncertainty RALF – explore or exploit Expected Error Reduction – reduce future error 35 Toward optimal active learning through sampling estimation of error reduction N. Roy and A. McCallum ICML 2001

36. 36 Expected Error Reduction

37. 2 Labeled Points 37 Ground Truth

38. Expected Error Reduction 38 Current Class Distribution Ground Truth

39. Expected Error Reduction 39 Compute the Expected Error (EE) for each unlabled datapoint Ground Truth

40. Expected Error Reduction 40 ? Hypothesize label 1 Ground Truth Class 1 Class 2

41. Expected Error Reduction 41 ? Update model Ground Truth

42. Expected Error Reduction 42 ? Hypothesize label 2 Ground Truth Class 1 Class 2

43. Expected Error Reduction 43 ? Update model Ground Truth

44. Expected Error Reduction 44 ? Compute EE Ground Truth

45. Expected Error Reduction 45 ? Hypothesize label 1 Ground Truth Class 1 Class 2

46. Expected Error Reduction 46 ? Update model Ground Truth

47. Expected Error Reduction 47 ? Hypothesize label 2 Ground Truth Class 1 Class 2

48. Expected Error Reduction 48 ? Update Model Ground Truth

49. Expected Error Reduction 49 ? Compute EE Ground Truth

50. Expected Error Reduction 50 Repeat for all unlabeled nodes! O(N 2 ) For Zhu et al. Ground Truth

51. Expected Error Reduction 51 Repeat for all unlabeled nodes! O(N 2 ) For Zhu et al. Ground Truth

52. Problems with EER Need to retrain the classifier with each unlabeled example (subquery) and for each different class label – O(N 2 ) At each step is it necessary to try every possible subquery? 52

53. 53 Active Learning Strategies Lower Complexity Performance RALF CVPR 2012 EER Zhu 2003 Random

54. Unsupervised Hierarchical Clustering 54

55. Unsupervised Hierarchical Clustering 55 Authority-shift clustering: Hierarchical clustering by authority seeking on graphs M. Cho and K. Mu Lee CVPR 2010 …

56. Unsupervised Hierarchical Clustering 56 …

57. Unsupervised Hierarchical Clustering 57 …

58. Unsupervised Hierarchical Clustering 58 Large clusters (exploration) Boundary refinement (exploitation) …

59. Our Hierarchical Subquery Evaluation After 2 Queries 59 Ground Truth

60. Our Hierarchical Subquery Evaluation 5.6 4.2 3.5 After 2 Queries Best EE Next nodes to add to the active set Current Active Set 60 Ground Truth Remaining Subqueries: 74

61. Our Hierarchical Subquery Evaluation Best EE After 2 Queries 61 Ground Truth 6 2.1 5.6 3.5 4.2 Remaining Subqueries: 2

62. Our Hierarchical Subquery Evaluation 6 2.1 3.2 1.1 After 2 Queries 62 Ground Truth 5.6 3.5 4.2 Remaining Subqueries: 0

63. Our Hierarchical Subquery Evaluation 6 2.1 After 3 Queries 3.2 1.1 Label for the example with the best EE is requested After 2 Queries 63 Ground Truth 5.6 3.5 4.2 Remaining Subqueries: 0

64. Our Hierarchical Subquery Evaluation After 3 Queries After 2 Queries 64 Ground Truth Remaining Subqueries: 72

65. 65 Results

66. 66 Results 1579 examples 8 classes 50 dim BoW PCA

67. 67 Results

68. 68 Results Ralf: A reinforced active learning formulation for object class recognition S. Ebert, M. Fritz, and B. Schiele CVPR 2012

69. 69 Results

70. 13 Different Computer Vision and Machine Learning Datasets 70 Results - Area Under Learning Curve

71. 13 Different Computer Vision and Machine Learning Datasets 71 Results - Area Under Learning Curve

72. Summary Hierarchical graph based semi-supervised active learning O(N 2 ) -> O(NlogN) 72

73. Summary Hierarchical graph based semi-supervised active learning O(N 2 ) -> O(NlogN) Robust to dataset type 73

74. Summary Hierarchical graph based semi-supervised active learning O(N 2 ) -> O(NlogN) Robust to dataset type Best user query in the time available 74

75. Future Work Representation learning – update graph structure during labeling 75

76. Representation learning – update graph structure during labeling Model different annotation costs 76 Future Work

77. Representation learning – update graph structure during labeling Model different annotation costs Embed new datapoints into the graph 77 Future Work

78. Thanks! 78 http://visual.cs.ucl.ac.uk/pubs/graphActiveLearning Come visit our poster 01-C-3

79. 79 Come visit our poster 01-C-3 http://visual.cs.ucl.ac.uk/pubs/graphActiveLearning

80. 80

81. 81 Graph Construction Comparison

82. 82 Timings

83. 83