Active Sites model for the B-Matrix Approach

TL;DR

This paper enhances the B-Matrix Hebbian learning approach by introducing active site identification during training, improving memory retrieval capacity through new update methods and comparative analysis.

Contribution

It proposes a novel active sites model that identifies key neurons during training to improve memory retrieval in B-Matrix based neural networks.

Findings

Active sites model improves memory retrieval accuracy.

New update methods outperform classical approaches.

Comparison shows the most effective method among proposed ones.

Abstract

This paper continues on the work of the B-Matrix approach in hebbian learning proposed by Dr. Kak. It reports the results on methods of improving the memory retrieval capacity of the hebbian neural network which implements the B-Matrix approach. Previously, the approach to retrieving the memories from the network was to clamp all the individual neurons separately and verify the integrity of these memories. Here we present a network with the capability to identify the "active sites" in the network during the training phase and use these "active sites" to generate the memories retrieved from these neurons. Three methods are proposed for obtaining the update order of the network from the proximity matrix when multiple neurons are to be clamped. We then present a comparison between the new methods to the classical case and also among the methods themselves.