Supervised perceptron learning vs unsupervised Hebbian unlearning: Approaching optimal memory retrieval in...

Area of Science:

• Computational neuroscience and artificial intelligence.
• Statistical mechanics and disordered systems.

Background:

• Hopfield-like neural networks utilize unsupervised learning for memory retrieval.
• Hebbian unlearning is a local algorithm aimed at enhancing network performance.
• Supervised learning algorithms, like training a linear symmetric perceptron, offer an alternative for network training.

Purpose of the Study:

• To numerically compare the Hebbian unlearning algorithm with a supervised training algorithm.
• To analyze the memory stability and learning dynamics of both approaches.
• To provide a geometric interpretation for the effectiveness of Hebbian unlearning.

Main Methods:

• Numerical comparison of Hebbian unlearning and supervised perceptron training.
• Analysis of basin of attraction sizes for stored memories.
• Investigation of convergence within Gardner's space of interactions.

Main Results:

• Hebbian unlearning yields basins of attraction comparable in size to those from supervised training.
• Both algorithms converge in similar regions of Gardner's interaction space, indicating parallel learning paths.
• A geometric interpretation is proposed to elucidate the optimal performance of Hebbian unlearning.

Conclusions:

• Hebbian unlearning is an effective unsupervised method for improving memory retrieval in neural networks.
• Its performance is comparable to supervised methods in terms of memory stability and learning convergence.
• Findings may extend to disordered magnetic systems and materials science for memory storage applications.