Inductive representation of conditional knowledge means to complete knowledge appropriately and can be looked upon as an instance of quite a general representation problem. The crucial problem of discovering relevant conditional relationships in statistical data can also be addressed in this formal framework. The main point in this paper is to consider knowledge discovery as an operation which is inverse to inductive knowledge representation, giving rise to phrasing the inverse representation problem. This allows us to embed knowledge discovery in a theoretical framework where the vague notion of relevance can be given a precise meaning: relevance heremeans relevancewith respect to an inductive representation method. In order to exemplify our ideas, we present an approach to compute sets of conditionals from statistical data, which are optimal with respect to the information-theoretical principle of maximum entropy.