Neural Noise

Multisynaptic spiking neuron for encoding spatiotemporal dynamics

2 min read

  • A multisynaptic spiking neuron model can simultaneously encode spatiotemporal dynamics (Nature).1
  • The model informs neural signal processing and neuromorphic or decoding models.1 1

Gardner updates

  • A multisynaptic spiking neuron model simultaneously encodes spatiotemporal dynamics, informing neural signal processing and neuromorphic or decoding models (Nature). 1

Weekly enrichment (2026-07-20)

  • The model is the Multi-Synaptic Firing (MSF) neuron, published in Nature Communications (2025) by Fan, Shen, Lian et al., in which a single axon can form multiple synapses with different firing thresholds on one postsynaptic neuron, jointly encoding spatial intensity via firing rates and temporal dynamics via spike timing.2
  • MSF neurons are shown to be a general abstraction that contains the standard Leaky Integrate-and-Fire (LIF) spiking neuron and the ReLU artificial-neuron activation as special cases, connecting spiking neural networks (SNNs) and conventional ANNs under one framework.2
  • The authors derive a theoretically optimal threshold-selection rule and a parameter-optimization criterion for surrogate-gradient functions under backpropagation-through-time (BPTT), which lets deep MSF-based SNNs scale up without the accuracy degradation seen in vanilla deep SNNs.2
  • On static image recognition with the MHSANet backbone, MSF (T=1, D=4) neurons matched or beat ReLU on CIFAR-10 (Top-1 97.28% vs 97.15%) while cutting energy 17.8x, and on ImageNet MSF (T=1, D=8) reached 79.89% vs ReLU’s 79.72% using 6.58 mJ vs 35.49 mJ (about 5.4x less power).2
  • On event-based/neuromorphic recognition, MSF improved accuracy over LIF by 1.4-2.0% on CIFAR10-DVS and also surpassed ReLU by 0.3-1.2%, demonstrating better spatiotemporal encoding on dynamic data.2
  • In signal-reconstruction tests, MSF-based spiking autoencoders achieved 29.6% lower image-reconstruction error (ImageNet) and 42.9% lower video-reconstruction error (HMDB51) than LIF-based counterparts.2
  • For object detection with a spiking YOLOX/MHSANet detector, MSF (T=1, D=4) gained 10.4% mAP over the LIF baseline on static COCO, and on the neuromorphic Gen1 dataset reached 72.4-73.7% mAP while using 16.6x less power than the ReLU version (5.5 mJ vs 91.5 mJ).2
  • The learned models reproduce a biological finding: multisynaptic connections are sparse (most neurons use a single synapse, with counts decaying roughly exponentially), matching synaptic-number distributions reported in a cubic millimeter of human temporal cortex.2
  • An official implementation of the MSF neuron and training scripts is released on GitHub, supporting reproducibility for neuromorphic and neural-decoding applications.3

Footnotes

  1. https://news.google.com/rss/articles/CBMiX0FVX3lxTE53RmU5Zy00aEF3Zmo2VXlzWWtWaXMxVkxFYmNnbldXV1BTd3BsSzVoUWxYeVc4amE0WlZ2a29VZ01WR1l3R0RRaUtESjRJQ0NfZzlSXzJPNXl2cml5dG5z?oc=5 ↩ ↩2 ↩3 ↩4

  2. https://www.nature.com/articles/s41467-025-62251-6 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  3. https://github.com/fanliangwei/Multisynaptic-spiking-neurons ↩

Graph View