A linear self-normalizing spike-timing-dependent plasticity rule for online learning in spiking neural networks

Spiking Neural Networks (SNNs) trained with Spike-Timing-Dependent Plasticity (STDP) are capable of continuous, online learning but critically suffer from unstable network dynamics. Homeostatic mechanisms such as divisive weight normalization improve stability to produce robust SNN models, however current normalization techniques are challenging to implement in neuromorphic architectures due to their computational complexity and memory access requirements. This work presents a novel, linear, self-normalizing STDP rule that drives convergence of each weight vector towards a target norm in an iterative and event-driven manner with O ( 1 ) space complexity. We achieve a mean normalization error of 1.5% for a target average weight of 0.10 on the MNIST and Fashion-MNIST datasets and improve training times on average by 37% compared to models trained with existing normalized STDP rules. The proposed approach leverages multiplicative weight dependence with a single scaling weight maximum term based on the L 0 norm of positive weight changes at each post-synaptic spike event. We also present a variant of the Self-Normalizing rule that learns both positive and negative weights while maintaining a mean of zero. This approach demonstrates impressive noise tolerance on inference tasks and opens avenues of development into the online training of kernels for convolutional networks. By incorporating the process of normalization into the event-driven STDP updates, the proposed Self-Normalizing rule offers a stable and efficient solution for neuromorphic applications where local STDP learning is supported without the need for computationally expensive weight normalization operations. • Synaptic weight normalization stabilizes spiking neural network dynamics. • Novel Self-Normalizing STDP rule steers synaptic weights to converge to a target mean. • Self-normalizing weight updates are local, linear, stable, and event-driven. • Weight mean of zero improves noise tolerance at inference. • Self-normalizing updates are more computationally efficient than existing techniques.