Invariant Object Representation Based on Principle of Maximum Dependence Capturing

Abstract

Sensory inputs conveying information about the environment are often noisy and incomplete, yet the brain can achieve remarkable consistency in object recognition. This cognitive robustness is thought to be enabled by transforming the varying input patterns into invariant representations of objects, but how this transformation occurs computationally remains unclear. Here we propose that sensory coding should follow a principle of maximal dependence capturing to encode associations among structural components that can uniquely identify objects. We show that a computational framework incorporating dimension expansion and a specific form of sparse coding can capture structures that contain maximum information about specific objects, allow redundancy coding, and enable consistent representation of object identities. Using symbol and face recognition, we demonstrate that a two-layer system can generate representations that remain invariant under conditions of occlusion, corruption, or high noise.