Neuroevosp Arse: A Biologically Plausible Framework for Efficient

SNNs have shown great potential in terms of power efficiency and event-driven processing during inference. To fully utilize their low power consumption and further improve their efficiency, researchers have explored pruning methods to create sparse SNNs without redundant connections after training. However, parameter redundancy still poses a challenge for SNNs during training. Inspired by the dynamic rewiring process observed in the human brain, we propose an efficient NeuroEvoSparse framework for SNNs, called NeuroEvoSparse-SNNs, to enable sparse SNN training from the beginning. In this framework, the pruning and regeneration of synaptic connections in SNNs evolve dynamically during learning, while maintaining a specific level of structural sparsity. This allows NeuroEvoSparse-SNNs to explore all potential parameters over time and find the optimal sparse connectivity. In summary, our proposed NeuroEvoSparse framework is a promising approach for training sparse SNNs from scratch. By incorporating biologically plausible evolutionary mechanisms and addressing the gap in expressiveness between sparse and dense training, this framework effectively learns SNNs with sparse structures while minimizing accuracy loss. This work has significant implications for lightweight training and inference of SNNs, which can lead to low power consumption and minimal memory usage in various applications.