About Breaking the Trade Off Between Accuracy and Comprehensibility in Concept Learning

The central issue of this paper is concerned with the knowledge representation used to encode inductively acquired concept descriptions. The central question being how to reconcile, in concept learning, the need for accurate representations (in terms of classification) as well as comprehensible ones (in terms of human-understandability)? It is known that language biases used in Machine Learning (ML) algorithms directly affect their performance as well as their comprehensibility. The problem is that, most of the time, the most "comprehensible" concept representations are not the best performer in terms of classification. In this paper, we argue that concept learning should be seen as a multiple-knowledge and multiple-functional model embedding a deep knowledge level optimized from recognition (classification task) and a shallow one optimized for comprehensibility (description task). This model assumes that the system has an interpretation function of the deep knowledge level that enables to build an "approximately correct" comprehensible description of it. We illustrate this approach with our GEM system which learns concepts in a numerical attribute space using a Neural Network representation at the deep knowledge level and symbolic decision rules at the shallow level.