Reproduce “DelGrad: Exact event-based gradients for training delays and weights on spiking neuromorphic hardware” 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
Julian Göltz, Jimmy Weber, Laura Kriener, Sebastian Billaudelle, Peter Lake, Johannes Schemmel, Melika Payvand, Mihai A. Petrovici
DelGrad enables exact, event-based backpropagation for jointly learning synaptic weights and continuous transmission delays in spiking neural networks. It improves classification over weight-only training in software and on BrainScaleS-2 hardware, reducing test error from 13.95% to 7.40% for a 30-neuron network with trained axonal delays.
Key insights
- Learned delays improve performance because they adapt the timing of incoming events to the task; they outperform fixed random delays, especially in smaller networks.
- Synaptic delays provide the greatest flexibility when networks are compared by neuron count, but axonal and dendritic delays offer more favorable parameter scaling and similar performance at comparable memory budgets.
- The usefulness of trainable delays depends on temporal scale: they are most effective when the available delay range can compensate for the timing relationships in the input data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.