In this paper, inspired by the complex dendritic structure of biological neurons, we propose NSPDI-SNN, an efficient and lightweight SNN method that incorporates nonlinear dendritic integration (NDI) and nonlinear synaptic pruning (NSP). NDI enhances the representation of spatiotemporal information in neurons, and NSP achieves high sparsity in SNNs. We conduct experiments on the DVS128 Gesture, CIFAR10-DVS, and CIFAR10 datasets, speech recognition, and reinforcement learning-based maze navigation tasks. In all tasks, we achieve high sparsity with minimal performance degradation. In particular, we achieve the best results on three event stream datasets, demonstrating that NSPDI significantly improves the efficiency of synaptic information transfer as sparsity increases. In conclusion, the complex structure of neuron dendrites and nonlinear computation demonstrate that NSPDI offers a promising approach for developing efficient SNN methods.
