Disordered Databases and Ordered Explanations
T. J. Parson · 1988
entropy, contextual infor-mation. 'Learning begins with organized knowledge which grows and becomes better organized '- Charnick and McDermont. The paper adopts an 'information theoretic' approach to an area of Machine Learning that is known as unsupervised, conceptual clustering. The approach is developed in the 'Vision Domain', databases of iipto 1000 entries being analysed con-taining information derived from digitised colour images. Typically, the field of supervised Machine Learning attempts to develop programs which learn from example and counter-example, failure or even instruction. Unsupervised Machine Learn ing algorithms adopt a different approach. The programs simply try to discover patterns, trends and hierachies of relationships which, although not explicitly stated, are never the less present within a database. In terms of the Vision problem, this corresponds to searching for significant classes or clusterings implicit in the attributes and relationships of im-age regions or features. The paper develops further the Knowledge rep-resentation structure and transformation rules presented at AVC87 [2]. A Hyperdigraph struc-ture for encoding relational knowledge is used, these being suitable for treatment from an in-formation theoretic viewpoint. By allowing non-disjoint classes to exist in this structure, it is shown that the transformation rules reduce the database to a minimal state of complexity. The uniqueness of this state allows classes of features, formed by the clustering process, to be ranked in importance by the increase in complexity within the hyperdigraph structure that their destruction would invoke. As regards implementation, the clustering pro-cess may be regarded as a "many-pattern / many-object " matching problem and the efficiency of such algorithms is discussed. Results based on the analysis of real images are presented. 1