From statistical inference to a differential learning rule for stochastic neural networks

Area of Science:

• Computational neuroscience
• Machine learning
• Artificial intelligence

Background:

• Stochastic neural networks model external stimuli probabilistically.
• Existing learning rules face challenges with biological feasibility and pattern storage.

Purpose of the Study:

• To derive a biologically plausible synaptic plasticity rule for stochastic neural networks.
• To enable efficient and extensive pattern storage while avoiding spurious attractors.

Main Methods:

• Derivation of a novel learning rule based on delayed activity correlations (delayed-correlations matching - DCM).
• Analysis of the rule's biological feasibility (finite signals, Dale's principle, locality).
• Evaluation of the rule's capacity for pattern storage in recurrent neural networks.

Main Results:

• The DCM rule demonstrates biological feasibility and supports extensive pattern storage.
• It handles correlated patterns, various architectures, and one-shot learning with the palimpsest property.
• The rule avoids the proliferation of spurious attractors.

Conclusions:

• The DCM rule offers a biologically plausible and effective mechanism for learning in stochastic neural networks.
• It facilitates the construction of generative models, like Boltzmann machines, for feature extraction and classification.