Bayesian retrieval in associative memories with storage errors.

Area of Science:

• Computational neuroscience
• Machine learning
• Statistical physics

Background:

• Single-step retrieval in autoassociative Willshaw networks is suboptimal for information extraction.
• Iterative retrieval strategies offer improved performance but lack theoretical grounding.

Purpose of the Study:

• To provide a principled, probabilistic inference framework for iterative retrieval in autoassociative networks.
• To develop and analyze novel iterative retrieval algorithms based on probabilistic principles.

Main Methods:

• Formulating retrieval as a probabilistic inference problem over exponentially many patterns.
• Developing two approximate, tractable iterative retrieval methods: maximum likelihood inference and mean field approximation.
• Analyzing the emergent properties of these methods, including Lyapunov functions and modified interaction terms.

Main Results:

• Iterative retrieval naturally emerges from probabilistic inference under noisy and corrupted conditions.
• Maximum likelihood inference yields a Lyapunov function for retrieval, incorporating an antiferromagnetic interaction term and site-dependent thresholds when storage errors are present.
• Mean field approximation leads to iterative equations interpretable as sigmoidal neural networks with similar interaction and threshold dynamics.

Conclusions:

• Probabilistic inference provides a strong theoretical justification for iterative retrieval strategies in autoassociative networks.
• The developed methods offer improved information retrieval from corrupted or noisy network states.
• The findings bridge concepts from machine learning, statistical physics, and neuroscience, offering new insights into neural network dynamics.