Next:
List of Figures
Up:
phd
Previous:
phd
Contents
List of Figures
List of Tables
Acknowledgements
Introduction
Summary of the thesis
Definition of The Problem, Terms Used and Basic Issues
Examples of Temporal Classification
Tech Support
Recognition of signs from a sign language
Definition of Terms
Channel
Stream
Frame
Stream Set
Statement of the problem
Weak temporal classification
Strong classification
Pre-segmented TC
Assessing success
Assessing Accuracy of Weak TC
Assessing Accuracy of Strong TC
Comprehensibility - A subjective goal
The major difficulties in temporal classification
Conclusion
Relationship to existing work
Related fields
Established techniques
Hidden Markov Models
Recurrent Neural Networks
Dynamic Time Warping
Recent interest from the AI community
Metafeatures: A Novel Feature Construction Technique
TClass
Overview
Tech Support revisited
Inspiration for metafeatures
Definition
Practical metafeatures
A more practical example
Representation
Extraction from a noisy signal
Using Metafeatures
Disparity Measures
Gain Ratio
Chi-Square test
Doing the search
Examples
With K-Means
With directed segmentation
Further refinements
Region membership measures
Bounds on variation
Conclusion
Building a Temporal Learner
Building a practical learner employing metafeatures
Expanding the scope of
TClass
Global attributes
Integration
Signal processing
Smoothing and filtering
Learners
Voting and Ensembles
Practical implementation
Architecture
Providing input
Implemented global extractors
Implemented segmenters
Implemented metafeatures
Developing metafeatures
Producing human-readable output in
TClass
Temporal and spatial Analysis of
TClass
Conclusion
Experimental Evaluation
Methodology
Practical details
Artificial datasets
Cylinder-Bell-Funnel - A warm-up
TTest
- An artificial dataset for temporal classification
Real-world datasets
Why these data sets?
Auslan
ECG
Conclusions
Future Work
Work on extending
TClass
Improving directed segmentation
Automatic metafeatures
Extending applications of metafeatures
Strong Temporal Classification
Speed and Space
Downsampling
Feature subset selection
Alternative approaches
Signal Matching
Approximate string-matching approaches
Inductive Logic Programming
Graph-based induction
Conclusions
Conclusion
Bibliography
Early versions of
TClass
Line-based segmentation
Per-class clustering
About this document ...
Mohammed Waleed Kadous 2002-12-10