Reinforcement Learning-Based Temporal Domain Regularization for Spiking Neural Networks

Spiking Neural Networks (SNNs), whose eventdriven accumulations replace the energy-hungry matrix multiplications of Artificial Neural Networks (ANNs), have become a prominent focus in neuromorphic research. Yet the recursive propagation of membrane potentials leaves them prone to temporal over-fitting, undermining generalization. Existing regularizers, largely transplanted from ANNs, target spatial domain and overlook SNN-specific time dynamics. We introduce a temporal regularization framework that stochastically retains or refreshes neuronal states at every time step, injecting noise directly along the temporal axis to curb over-fitting at its source. The framework comes in two variants: a lightweight fixed-probability version that is inexpensive to implement, and an adaptive reinforcement-learning version that offers finer control. Evaluations on several neuromorphic benchmarks confirm that both variants consistently enhance accuracy and robustness.