Freezing transitions in neural networks

Abstract

I consider neural networks as feedback dynamical systems for retrieving information, which in the retrieval phase is driven to an attractor correlated with a stored pattern. Dynamics in this attractor may be described by considering the activity distribution. This enables us to determine the degree of chaoticity of the network dynamics. Consequently I am able to demonstrate that as more patterns are stored the system becomes more chaotic, and undergoes a transition for a partially frozen phase to an unforzen phase. Improvement in retrieval using a freezing procedure is also discussed.