Sifting through a large hypothesis space: Revisiting differentiable learning through satisfiability ∗

Abstract

A difficultly which must be addressed by inductive logical programming (ILP) systems is how to deal with the enormous space of plausible solutions. The majority of modern ILP systems approach this problem through the meta-learning paradigm, that is, only consider plausible solutions which are constructable from a set of clause templates. This approach has been adopted by investigations into neuro-symbolic ILP. Our investigation uses clause templates together with a variant of δILP, to expand the hypothesis space, rather than contract it. Our experiments support the following hypothesis: providing gradient descent with a larger solution space aids the discovery of explanatory hypotheses.