List of Figures

1. The gloss for the Auslan sign thank [Joh89].
2. An example of a stream from the sign recognition domain with the label come.
3. The relationship between channels, frames and streams.
4. A diagram illustrating an HMM and the different ways a,a,b,c can be generated by the HMM.
5. A typical feedforward neural network.
6. Recurrent neural network architecture.
7. Dynamic time warping.
8. Parameter space for the LoudRun metafeature applied to the Tech Support domain.
9. Parameter space for the LoudRun metafeature in the Tech Support domain, but this time showing class information. Note that the point (3,3) is a ``double-up''.
10. Three synthetic events and the regions around them for LoudRuns in the Tech Support domain.
11. Rule for telling happy and angry customers apart.
12. The gloss for the sign building, from [Joh89].
13. An example of the variation in the y value of the sign building.
14. Two different possible representations of an Increasing event.
15. The y channel of another instance of the sign building, but this one with much greater noise than Figure 4.6.
16. Instantiated features extracted from Figure 4.6.
17. Instantiated features extracted from Figure 4.8.
18. K-means algorithm.
19. Random search algorithm.
20. K-means clustered LoudRun parameter space for Tech Support domain.
21. Trial 1 points and region boundaries.
22. Trial 2 points and region boundaries.
23. Trial 3 points and region boundaries.
24. Rule for telling happy and angry customers apart, using synthetic features from trial 2.
25. Human readable form of rules in Figure 4.17.
26. Human readable form of rules in Figure 4.4.
27. Distance measures in the parameter space. $c_1$ and $c_2$ are two centroids, A and B are instantiated features.
28. Rule for telling happy and angry customers apart, using synthetic features from trial 2 and using relative membership.
29. The bounding boxes in the case of Trial 2 random search.
30. Human readable form of rules in Figure 4.17.
31. Transforming relative membership into readable rules.
32. The stages in the application of a single metafeature in TClass.
33. The TClass pipeline for processing test instances.
34. Instantiated feature extraction.
35. Synthetic feature construction.
36. Training set attribution.
37. The TClass system: training stage.
38. The TClass system: testing stage.
39. Domain description file for Tech Support Domain.
40. Domain description file for Powerglove Auslan Domain.
41. An example of a TClass Stream Data (tsd) file.
42. An example of a TClass Stream Data (tsd) file from the sign language domain.
43. An example of a TClass class label file from the Tech Support domain.
44. Component description file for Tech Support domain.
45. Learnt classifier for Tech Support domain after running TClass on it.
46. Post-processing Figure 5.14 to make it more readable.
47. Cylinder-bell-funnel examples.
48. An instance of the trees produced by naive segmentation for the CBF domain.
49. One decision tree produced by TClass on the CBF domain.
50. Events used by the decision tree in Figure 6.3.
51. A ruleset produced by TClass using PART as the learner.
52. Event index for the ruleset in Figure 6.5.
53. Prototype for class A
54. Prototype for class B
55. Prototype for class C
56. Effect of adding duration variation to prototypes of class A.
57. Effect of adding Gaussian noise to prototypes of class A.
58. Effect of adding sub-event variation to prototypes of class A.
59. Effect of adding amplitude variation to prototypes of class A.
60. Effect of replacing gamma channel with irrelevant signal to class A.
61. Examples of class A with default parameters.
62. Examples of class B with default parameters.
63. Examples of class C with default parameters.
64. Learner accuracy and noise
65. Voting different runs of TClass to reduce error with g=0.2.
66. Error rates of different learners with 100 and 1000 examples.
67. A decision tree produced by TClass on TTest with no noise. It is a perfect answer; ignoring gamma altogether as a noisy channel and having the absolute minimum number of nodes.
68. Events used by the decision tree in Figure 6.21.
69. A decision list produced by TClass on TTest with 10 per cent noise.
70. Events used by the decision tree in Figure 6.23.
71. A decision tree produced by TClass on TTest with 20 per cent noise.
72. Voting TClass generated classifiers approaches the error of hand-selected features.
73. A decision list produced by TClass on the Nintendo sign data for the sign thank.
74. Events referred to in Figure 6.27 with bounds.
75. Effect of voting on the Flock sign data domain.
76. A decision list produced by TClass on the Flock sign data for the sign thank.
77. Events referred to in Figure 6.27 with bounds.
78. The components of a heartbeat. Taken from [dC98].
79. The ECG data as it arrives when seen by TClass.
80. The effect of applying de Chazal's filter to the ECG data (taken from [dC98]).
81. Voting TClass learners improves accuracy.
82. Voting TClass learners asymptotes to error the hand-selected feature extraction technique.
83. A two way classifier for the RVH class in the ECG domain.
84. Events referenced by Figure 6.37.
85. Random search algorithm, using the learner itself
86. The SMI of the z channel of one instance of building with itself.
87. The SMI of the z channel of an instance of building and an instance of make.
88. The subword tree for abracadabra.
89. Line-based segmentation example for y channel of sign come.
90. The parameter space of local maxima of the y channel in the Flock sign domain. All instantiated local maxima are shown.
91. The parameter space of local maxima of the y channel in the Flock sign domain, but shown for only two classes.