Neural network approach to relevance

It is becoming increasingly recognised in AI and elsewhere that it is important to understand and account for relevance for computational and functional reasons. In this paper the authors give an account of relevance formally, and present a neural network approach to relevance with a focus on how to preserve relevant information and ignore irrelevant information in order to facilitate decision making and thereby to improve performance. This can provide a basis for further discussion of relevance as well as having immediate significance for neural networks. The authors' approach is based on a novel associative network model, Moving Around Landscape (MAL), which is based informally on kinetics and is concerned mainly with information transition under constraints. MAL facilitates accommodating relevance in its dynamics to evolve computations that ignore irrelevant aspects of the environments and preserve relevant information. Experiments on its application to character recognition show that accommodating one kind of relevance (neighbourhood relationship) into MAL enables it to learn relevant internal representations of character images that are comparable in terms of ordinary metrics, such as Hamming distance and Euclidean distance. Thereby character recognition with physical variations is facilitated and the performance of recognition may be improved.