INDEPENDENCE OR NON-NEGATIVITY: CONSTRAINTS FOR SENSORY CODING

Abstract

This paper addresses two different coding strategies in the framework of linear generative models. One relies on statistical independence between encoding variables and the other imposes non-negativity constraints in both basis functions and encoding variables. These lead to holistic or parts-based representation. We stress out some useful behaviour of parts-based representation and introduce a new way of information representation, which exhibits holistic structure as well as parts-based representation together.