Multiple-knowledge representations in concept learning

This paper investigates a general framework for learning concepts that allows to generate accurate and comprehensible concept representations. It is known that biases used in learning algorithms directly affect their performance as well as their comprehensibility. A critical problem is that, most of the time, the most ''comprehensible'' representations are not the best performer in terms of classification! In this paper, we argue that concept learning systems should employ Multiple-Knowledge Representation: a deep knowledge level optimised from recognition (classification task) and a shallow one optimised for comprehensibility (description task). Such a model of concept learning assumes that the system can use an interpretation function of the deep knowledge level to build an approximately correct comprehensible description of it. This approach is illustrated through our GEM system which learns concepts in a numerical attribute space using a Neural Network representation as the deep knowledge level and symbolic rules as the shallow level.