Generative Conceptual Representations and Semantic Communications

Abstract

Representations are essential in learning of natural and artificial systems due to their ability to identify characteristic patterns in the sensory inputs. In this work we examined latent representations of images of basic geometric shapes and handwritten digits as a basis for sharing semantic information about observations in a collective of unsupervised generative learners. Individual models trained in an unsupervised process with minimization of generative error were exposed to a process of synchronization of symbolic tokens associated with characteristic regions in the latent representations identified with two different strategies. It was demonstrated that conceptual representations with good decoupling of characteristic patterns can be produced reliably and consistently with models of unsupervised generative self-learning; and that a simple process of conceptual synchronization can enable effective sharing of information between individuals in a collective by associating shared symbols with latent regions correlated with characteristic patterns in the sensory inputs. The results demonstrate the potential of conceptual latent representations as a natural platform for development of abstract concept intelligence and communications.