A Bayesian Framework to Integrate Symbolic and Neural Learning

Abstract

In this paper we show how a Bayesian framework can be used to get a unified description of a variety of learning methods integrating symbolic and neural learning. The constraints supplied by symbolic component of a hybrid learning method can be viewed as Bayesian priors. In this perspective, neural component can be described using usual Bayesian formalism plus taking into account the derived priors. There are several advantages of the proposed formalism. First, we get a method for designing appropriate learning rules and adjusting their parameters. Second, we obtain the possibility to compare different hybrid learning methods through evaluating their evidence. Third, Bayesian formalism may lead to a new probabilistic measure for knowledge refinement. We will demonstrate how to use in practice the proposed formalism by constructing on its basis a neural method learning symbolic knowledge. Experiments on the hard problem of the protein secondary structure prediction validate this method.