Reproduce “Exploring Tradeoffs 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
Florian Bacho, Dominique Chu
This paper investigates tradeoffs in Spiking Neural Networks by proposing a multi-spike extension to the Time-To-First-Spike (TTFS) constrained Fast & Deep algorithm. The research demonstrates that allowing neurons to fire multiple times leads to improved accuracy, enhanced robustness to noise and weight quantization, and faster convergence, while maintaining comparable sparsity and competitive prediction latency across image and temporal datasets.
Key insights
- Strict Time-To-First-Spike (TTFS) constraints in SNNs, while promoting speed, limit performance and robustness, especially for temporal data.
- Allowing multiple spikes per neuron introduces redundancy and enables the network to leverage the full temporal extent of inputs, leading to improved performance and resilience.
- Sparsity in multi-spike SNNs can emerge naturally through learning, where error propagation due to output spikes influences hidden layer activity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.