Cognitively inspired rule-plus-exemplar based pattern classification

• Main Author: Sit, Wing Yee
• Other Authors: Mao Kezhi
• Format: Theses and Dissertations
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Pattern recognition
• Online Access: https://hdl.handle.net/10356/61749
• Institution: Nanyang Technological University

Despite the many strengths of machine learning pattern classification techniques, they have intrinsic weaknesses compared to human learning, which is far more comprehensible, adaptive, and flexible. These are issues that need to be addressed. In applications where the classifiers play supporting roles in decision making, comprehensibility is required, restricting the kinds of classifiers that can be used. In a dynamic environment, adaptation is important, particularly in view of evolving concepts. The training data that was originally provided may not be sufficient to represent the concept as it evolves. These qualities natural to humans are significantly more difficult to achieve efficiently in machines. To imitate these desired qualities, ideas are drawn from cognitive psychology, based on which machine learning classifiers are structured into systems that can better deal with such tasks. The relevant categorization models using rules and exemplars are employed as the basis of such systems. First, we explore the comprehensibility issue. This trait is most directly achieved through the use of explicit IF-THEN rules that are directly interpretable to humans. However, comprehensibility often comes at reduced complexity of the classifier and thus poorer generalization performance. This tradeoff between interpretability and accuracy is well recognized. As humans use both the rule and exemplar models in cognitive psychology for categorization tasks, these models are intuitive to us and are largely sufficient in most tasks. The use of the exemplar model thus complements and aids the commonly used rule model in classification to improve generalization capability. The resulting system shows marked improvement in classification accuracy compared to common interpretable classifiers. The next problem considered is that of extrapolative generalizability. Classifiers learn from training data that is assumed to be representative of the concept to be learnt, but this condition is not always satisfied. Test samples sometimes do not arise from the represented concept, resulting in classification that is inconsistent and erratic. We propose a framework for dealing with such data, using the rule and exemplar models from cognitive psychology. The key to such a framework lies in the correct identification of the training region, which is the region in input feature space that has been covered by the training samples. The method provided is non-parametric and not dependent on any specific classification technique used. There are also no assumptions on the shape and density of the training samples in the input feature space. Experiments show that the system can extrapolate well as the samples outside the training region are correctly identified and appropriately handled. Adaptability of classifiers is another desired quality but has received far more attention than the previous two problems. Changing concepts have been explored in the context of concept drift. Due to the change in concept, the classifier needs to adapt quickly so that it remains relevant and can properly classify the sample stream that is being presented to it. However, the problem of concept growth is also present in many applications. It is more difficult to solve because unlike in concept drift, the previous data cannot simply be forgotten. The rule-plus-exemplar structure is adopted in a new system that specifically deals with concept growth, and results show that the system achieves a good balance of stability and plasticity to learn and correctly classify a changing sample stream. Finally, as the concept growth problem often occurs in applications with imbalanced classes, a different learning and classification system is given. Even though it is based on the rule-plus-exemplar structure, the models are incorporated in a completely different way to deal with both issues simultaneously. There has been little existing work on this problem, and the available methods are compared with the proposed system. The performance of the rule-plus-exemplar system is significantly better and more consistent than other methods.