Proceedings of the Annual Meeting of the Cognitive Science Society

Human conceptual knowledge supports the ability to generatenovel yet highly structured concepts, and the form of this con-ceptual knowledge is of great interest to cognitive scientists.One tradition has emphasized structured knowledge, view-ing concepts as embedded in intuitive theories or organizedin complex symbolic knowledge structures. A second tradi-tion has emphasized statistical knowledge, viewing conceptualknowledge as an emerging from the rich correlational structurecaptured by training neural networks and other statistical mod-els. In this paper, we explore a synthesis of these two traditionsthrough a novel neuro-symbolic model for generating new con-cepts. Using simple visual concepts as a testbed, we bring to-gether neural networks and symbolic probabilistic programsto learn a generative model of novel handwritten characters.Two alternative models are explored with more generic neuralnetwork architectures. We compare each of these three mod-els for their likelihoods on held-out character classes and forthe quality of their productions, finding that our hybrid modellearns the most convincing representation and generalizes fur-ther from the training observations.