Storage capacity of two-dimensional neural networks

Area of Science:

• Statistical mechanics
• Computational neuroscience
• Artificial neural networks

Background:

• The Hopfield model is a recurrent neural network used for associative memory.
• Understanding the capacity of these networks to store and retrieve patterns is crucial for their application.
• Previous studies have explored phase diagrams and retrieval properties.

Purpose of the Study:

• To determine the maximum number of patterns that can be reliably embedded and retrieved in a two-dimensional Hopfield network.
• To analyze the energy landscape and stability of network states related to pattern retrieval.
• To provide a theoretical limit on the storage capacity of this model.

Main Methods:

• Exact calculation of grand state energies for specific network configurations.
• Utilizing the correlation function of the ferromagnetic Ising model.
• Investigating the energy landscape and stability of the pure retrieval state.
• Referencing established phase diagrams from prior research.

Main Results:

• The grand state energies of the pure-ferromagnetic state and a single stored pattern were precisely calculated.
• Analysis of the energy landscape revealed insights into the stability of retrieved patterns.
• The study confirmed limitations on the number of retrievable patterns.

Conclusions:

• The two-dimensional Hopfield network is limited to retrieving a maximum of three embedded patterns.
• The findings are consistent with qualitative features of known phase diagrams for such networks.
• This provides a concrete upper bound for associative memory capacity in this model.