TR-SNN: a lightweight spiking neural network based on tensor ring decomposition

Aim Spiking neural networks (SNNs), inspired by biological neural mechanisms, have demonstrated remarkable capabilities in various intelligent tasks. However, addressing more complex scenarios often requires larger SNN models with a vast number of parameters, leading to challenges in model training and hardware deployment. To mitigate these issues, researchers have developed various lightweight strategies for compressing SNN model parameters, primarily categorized into pruning, quantization, and distillation. Nevertheless, these methods often involve complex training processes and are prone to significant accuracy loss.Methods In this work, we propose a novel TR-SNN model that achieves parameter compression using tensor ring decomposition. By expressing weight tensors as a sequence of elementary operations acting on other, TR-SNN models significantly reduce parameter complexity.Results Extensive experiments demonstrate that TR-SNN models achieve competitive performance compared to other state-of-the-art methods, effectively reducing the number of parameters while maintaining accuracy.Conlusion Overall, our findings highlight the potential of tensor ring decomposition as a promising approach for developing lightweight SNN models.