Active perception based on free-energy minimization on restricted Boltzmann machines

Abstract

A robot, which would like to respond quickly in the world, should select more informative signals to estimate the cause of its sensation (e.g., a state of the environment, a category of a handling object, an emotional state of interaction partner, etc.). This paper proposes an active perception framework that selects the robot's action to perceive critical sensory signals based on a free-energy minimization in an energy-based model. We employed a restricted Boltzmann machine as a fundamental component for an estimation network of the cause of sensations. Our framework demonstrated better performance for the attention control in emotional human-robot interaction than other methods.