Capsule Generative Models

Neuroscience studies inspire that structures are needed in the hidden space of deep learning models. In this paper, we propose a capsule restricted Boltzmann machine and a capsule Helmholtz machine by replacing individual hidden variables with encapsulated groups of hidden variables. Our preliminary experiments show that capsule activities in both models can be dynamically determined in context, and these activity spectra exhibit between-class patterns and within-class variations. Our models offer a novel approach to visualizing and understanding the hidden states.