Brain-inspired reward broadcasting: Brain learning mechanism guides learning of spiking neural network

The secret of the brain's efficient learning lies in its rich set of learning mechanisms, neuronal encoding and plasticity principles. However, the combination of learning mechanisms and plasticity principles has not yet been incorporated into spiking neural networks (SNN). Existing brain-inspired learning algorithms still have problems such as being unable to achieve global credit allocation and causing catastrophic forgetting in incremental learning. Drawing inspiration from the brain's reward learning mechanism, this paper employs reinforcement learning (RL) to model reward mechanisms and achieve synaptic plasticity and presents a novel SNN learning algorithm called Brain-Inspired Reward Broadcasting Algorithm (BIRB). BIRB treats synapses as agents and guides agent actions to achieve plasticity through RL; it is biologically plausible to model plasticity in this way as it achieves global credit assignment by broadcasting reward signals, which aligns with the reward mechanism process. BIRB incorporates a innovative learning approach known as two-step learning, which effectively addresses the issue of catastrophic forgetting. This paper also mathematically derives the plasticity process and learning principles of BIRB. Moreover, experimental results demonstrate the exceptional performance of BIRB across multiple datasets, thereby highlighting its pivotal role in guiding synaptic weight changes and facilitating generalization.