Two-factor synaptic consolidation reconciles robustness with pruning and homeostatic scaling

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Memory consolidation stabilizes neural representations, primarily during sleep.
• Existing computational models struggle to integrate consolidation mechanisms like neural replay and synaptic plasticity.
• Understanding these mechanisms is key to developing robust learning and memory systems.

Purpose of the Study:

• To propose a unified mathematical model for memory consolidation.
• To explain how neural replay and synaptic plasticity contribute to memory stabilization.
• To account for experimentally observed consolidation phenomena within a computational framework.

Main Methods:

• Derived a self-supervised consolidation model incorporating neural replay and two-factor synapses.
• Analyzed the model's dynamics to understand memory encoding and robustness to synaptic noise.
• Compared model predictions with experimental data on synaptic scaling, pruning, and selectivity.

Main Results:

• The model demonstrates how replay and synaptic mechanisms lead to sparse connectivity and robust cued recall.
• It unifies multiplicative homeostatic scaling, task-driven synaptic pruning, and increased neural selectivity.
• The model accurately reproduces developmental trends and predicts sublinear scaling of synaptic noise with strength.

Conclusions:

• The proposed model offers a unified computational account of memory consolidation.
• It highlights the role of synaptic noise in shaping memory network properties.
• Findings support the model's ability to explain diverse consolidation phenomena and developmental trajectories.