Coherent and Consistent Relational Transfer Learning with Auto-encoders

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 and properties. That is, they should be consistent with respect to relational constraints, and this consistency must extend beyond the representations encountered in the source domain. Further, where relations are modelled by differentiable functions, their gradients must conform - the functions must at times move together to preserve consistency. We propose a Partial Relation Transfer (PRT) task which exposes how well relation-decoders model these properties, and exemplify this with ordinality prediction transfer task, including a new data set for the transfer domain. We evaluate this on existing relation-decoder models, as well as a novel model designed around the principles of consistency and gradient conformity. Results show that consistency across broad regions of input space indicates good transfer performance, and that good gradient conformity facilitates consistency.