Reproduce “Neural Architecture Search for 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

Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, Priyadarshini Panda

Researchers at Yale University developed an efficient Neural Architecture Search (NAS) method tailored for Spiking Neural Networks (SNNs), discovering architectures with novel temporal feedback connections. These automatically designed SNNs achieved new state-of-the-art accuracy on image recognition tasks while operating at significantly lower timesteps than previous SNN models.

Key insights

  • SNNs achieve superior performance when employing architectures specifically tailored to their temporal dynamics, particularly those that incorporate backward (temporal feedback) connections.
  • The Sparsity-Aware Hamming Distance (SAHD) is crucial for accurately assessing SNN architectures at initialization, providing a more reliable correlation with post-training accuracy than standard Hamming Distance.
  • Automatically searched SNN architectures, especially those with backward connections, can achieve state-of-the-art accuracy with significantly fewer timesteps, indicating high energy efficiency and low-latency potential.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.