Reproduce “Beyond Rate Coding: Surrogate Gradients Enable Spike Timing Learning in Spiking Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ziqiao Yu, Pengfei Sun, Danyal Akarca, Dan F. M. Goodman
The authors test whether surrogate-gradient-trained spiking networks use timing rather than counts, using controlled tests and speech-derived data equalized to each neuron's count across samples. They compare networks with and without trainable axonal delays. On SHD-Norm, accuracy is 48% with delays and 23% without, versus the 5% count-only chance baseline. Because subsampling can distort timing, results show retained timing supports classification, not preservation of every original timing feature.
Key insights
- Surrogate gradient descent algorithms enable SNNs to genuinely learn and exploit various forms of precise spike timing (ISIs, CCISIs, coincidence codes), extending their capabilities beyond simple firing rate computations.
- Learnable axonal delays play a critical and complementary role, significantly enhancing SNN performance on tasks involving longer temporal dependencies (e.g., cross-channel ISIs) and in real-world datasets where rate information is controlled, by providing mechanisms for precise temporal alignment.
- SNNs augmented with learnable delays develop a heightened sensitivity to the temporal order and causal structure of input spike patterns, aligning more closely with how biological systems perceive sequential information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.