Reproduce “Error-triggered Three-Factor Learning Dynamics for Crossbar Arrays” 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

Melika Payvand, Mohammed Fouda, Fadi Kurdahi, Ahmed Eltawil, Emre O. Neftci

A subthreshold analog circuit architecture is introduced for crossbar arrays that implements an error-triggered, asynchronous Surrogate Gradient (SG) learning rule for Spiking Neural Networks. This approach achieves up to a hundred-fold reduction in synaptic weight updates with minimal impact on recognition accuracy for event-based datasets like N-MNIST and DVS Gestures.

Key insights

  • Sparsity in weight updates, achieved through an error-triggered mechanism, can drastically reduce energy consumption and extend device lifetime without significantly compromising learning performance.
  • The same physical circuit components can be effectively shared between inference and training dynamics, simplifying hardware design and enhancing robustness to fabrication mismatch.
  • Analog subthreshold circuits provide a viable and efficient pathway for implementing complex Surrogate Gradient learning rules directly on-chip within memristor crossbar arrays.

Summary by alphaXiv.

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