Neural structures to compute homothetic invariances for artificial perception systems

Abstract

This paper is oriented to a computational theory of invariant perception by the cortex.

Based on the idea that the cortex has adopted representations and computational strategies that make the computation of invariants efficient, we suggest that in the cortex there are, at least, two paths for computing invariances. A path computes the parameters of invariance and the other applies them to original sensorial patterns. The neuronal structures that we propose in this paper uphold both the architecture and functionality of the cortex.

We present a model of neural net which computes the homothetic parameter of an one-dimensional tonotopic pattern. This theoretical problem has, as substratum, the spatially layered architecture of Primary Auditory Cortex, (AI), and the associated computational concepts. The neuronal synthesis of the model is achieved combining McCulloch-Pitts and analytical formulations, which allows us to obtain a neural layered computing structure.

In addition, we propose an alternative net for computing invariances which is derived from methods of artificial systems visual processing. These neural structures, when working in two-dimensional spaces, allow reaching, in a natural way, a complete schema of recognition for artificial visual systems.