Learning with a network of Competing Synapses

Abstract

Competition between synapses arises in correlation-based plasticity. We propose a game-theory-inspired model in which synapses switch between weak and strong states, driven by local competition and memory of past outcomes. The framework reproduces power-law forgetting and captures motor adaptation phenomena like savings and rebound via distinct timescales.

Methodology

We model each synapse as a binary unit updated by parallel/sequential rules. Synaptic transition probabilities depend on neighbors’ states and a memory parameter. We derive mean-field equations, identify phases via simulations, and compare system-level learning and forgetting curves to existing motor adaptation data. Competition between synapses arises in correlation-based plasticity. We propose a game-theory-inspired model in which synapses switch between weak and strong states, driven by local competition and memory of past outcomes. The framework reproduces power-law forgetting and captures motor adaptation phenomena like savings and rebound via distinct timescales.

We model each synapse as a binary unit updated by parallel/sequential rules. Synaptic transition probabilities depend on neighbors’ states and a memory parameter. We derive mean-field equations, identify phases via simulations, and compare system-level learning and forgetting curves to existing motor adaptation data. Competition between synapses arises in correlation-based plasticity. We propose a game-theory-inspired model in which synapses switch between weak and strong states, driven by local competition and memory of past outcomes. The framework reproduces power-law forgetting and captures motor adaptation phenomena like savings and rebound via distinct timescales.