Reproduce “On-Chip Error-triggered Learning of Multi-layer Memristive 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
Melika Payvand, Mohammed E. Fouda, Fadi Kurdahi, Ahmed M. Eltawil, Emre O. Neftci
An on-chip error-triggered learning algorithm and corresponding neuromorphic hardware architecture are presented for multi-layer memristive Spiking Neural Networks, achieving up to 88 times fewer memristor write operations. This approach maintains competitive classification accuracy, with error rates of 5.36% on DVS Gestures and 3.00% on N-MNIST, suitable for online, energy-efficient learning.
Key insights
- Factorizing the gradient-based learning rule into intrinsic (local in NCs) and extrinsic (computed by PCs) components enables scalable, distributed on-chip training for multi-layer SNNs without requiring global backpropagation.
- An error-triggered mechanism significantly reduces memristor write operations by only updating weights when a local error exceeds a tunable threshold, leading to substantial energy savings and extended device lifetime.
- Despite using hardware-friendly approximations like binarized post-synaptic potentials and ternary weight updates, the algorithm maintains competitive classification accuracy, demonstrating a viable trade-off for on-chip efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.