Reproduce “Meta-learning Spiking Neural Networks with Surrogate Gradient Descent” 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

Kenneth Stewart, Emre Neftci

Researchers from UC Irvine, Forschungszentrum Jülich, and RWTH Aachen present a method to apply second-order meta-learning to Spiking Neural Networks (SNNs) by utilizing twice-differentiable surrogate gradients. This approach enables SNNs to achieve rapid, few-shot adaptation on event-based data, demonstrating accuracy comparable to conventional ANNs and producing large-magnitude weight updates suitable for low-precision neuromorphic hardware.

Key insights

  • Surrogate gradient methods for SNNs can be designed to be twice differentiable, enabling the application of advanced second-order meta-learning algorithms like MAML.
  • Meta-learning allows SNNs to acquire an optimal initialization that facilitates rapid adaptation to new tasks with minimal data and few inner-loop updates, compatible with neuromorphic hardware constraints.
  • The meta-learned SNNs generate large-magnitude weight updates during adaptation, which is beneficial for low-precision memory often found in neuromorphic computing devices.

Summary by alphaXiv.

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