Coherent and Consistent Relational Transfer Learning with Autoencoders

Human defined concepts are inherently transferable, but it is not clear under what conditions they can be modelled effectively by non-symbolic artificial learners. This paper argues that for a transferable concept to be learned, the system of relations that define it must be coherent across domains. This is to say that the learned concept-specific relations ought to be consistent with respect to a theory that constrains their semantics and that such consistency must extend beyond the representations encountered in the source domain. To demonstrate this, we first present formal definitions for consistency and coherence, and a proposed Dynamic Comparator relation-decoder model designed around these principles. We then perform a proposed Partial Relation Transfer learning task on a novel data set, using a neural-symbolic autoencoder architecture that combines sub-symbolic representations with modular relation-decoders. By comparing against several existing relation-decoder models, our experiments show that relation-decoders which maintain consistency over unobserved regions of representational space retain coherence across domains, whilst achieving better transfer learning performance.