The paper introduces the Multi-Synaptic Firing (MSF) neuron, a novel computational model for spiking neural networks (SNNs) inspired by biological multisynaptic connections. Unlike traditional models that use a single threshold, MSF neurons utilize multiple firing thresholds to simultaneously encode spatial intensity through firing rates and temporal dynamics through precise spike timing. This architecture effectively generalizes existing LIF and ReLU neurons, bridging the performance gap between artificial and spiking neural paradigms while maintaining low power consumption and latency. To support deep network scalability, the authors derive optimal threshold selection and parameter criteria that prevent gradient degradation during training. Extensive experiments across benchmarks—including object detection, biological signal processing, and reinforcement learning—demonstrate that MSF-based SNNs significantly outperform conventional models in handling complex spatiotemporal tasks. Ultimately, this research mimics the sparse, strong connections found in the human cerebral cortex to advance the efficiency of real-world neuromorphic computing.

References:

• Fan L, Shen H, Lian X, et al. A multisynaptic spiking neuron for simultaneously encoding spatiotemporal dynamics[J]. Nature Communications, 2025, 16(1): 7155.