q-state Potts-glass neural network based on pseudoinverse rule

Area of Science:

• Computational neuroscience
• Statistical physics

Background:

• The q-state Potts-glass neural network is a model for associative memory.
• Traditional Hebbian learning rules can lead to spurious memories and metastable states.

Purpose of the Study:

• To investigate the performance of the q-state Potts-glass neural network using the pseudoinverse (PI) rule.
• To compare the PI rule with the Hebbian rule.
• To identify novel dynamical phases and their properties.

Main Methods:

• Simulations of the q-state Potts-glass neural network with both PI and Hebbian learning rules.
• Analysis of storage capacity, retrieval quality, and network dynamics.
• Identification of critical points and dynamical phases.

Main Results:

• The PI rule improves storage capacity and retrieval quality for q<14.
• Network dynamics differ significantly between PI and Hebbian rules.
• A novel dynamical phase, observed for q≥3, completely suppresses spurious memories, irrespective of the learning rule.

Conclusions:

• The PI rule offers a significant advantage for Potts-glass neural networks under certain conditions.
• The discovered dynamical phase eliminates spurious memories, preventing entrapment in metastable states.
• This finding enhances the potential applications of multistate Potts-glass neural networks.