Synaptic Weight Optimization for Oscillatory Neural Networks: A Multi-Agent RL Approach

The Oscillatory Neural Network (ONN) presents itself as a promising architecture model for pattern recognition (PR), based on which advanced neuromorphic computing and integrated circuit designs are implemented. The core of the ONN’s PR capability lies in the synaptic weight design, i.e., how the neurons are connected to each other. Conventional design methods, like the Hebbian rule, are able to store only a limited number of patterns. In this paper, we propose a strategy to leverage the Multi-Agent Reinforcement Learning (MARL) for acquiring the optimal synaptic weights that can efficiently store more patterns into the ONN system as stable quilibria. To obtain the synaptic weights in a more efficient manner and further increase the number of patterns to be stored, we additionally propose a method to leverage Curriculum Learning (CL) to optimize the learning process of the policy. Experimental results demonstrate that the proposed MARL-based method outperforms baseline methods in terms of storing more patterns as stable equilibria in ONN.