Reproduce “Spiking Deep Networks with LIF Neurons” 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

Eric Hunsberger, Chris Eliasmith

A backpropagation-compatible method enables deep convolutional networks built from leaky integrate-and-fire neurons by smoothing the LIF firing-rate function during training and injecting noise to model spike-train variability. The converted networks achieve 17.05% error on CIFAR-10 and 1.63% on MNIST, improving over earlier reported LIF and integrate-and-fire results while retaining biologically motivated neuron dynamics.

Key insights

  • Moderate noise improves rate-to-spike conversion robustness, but excessive noise degrades both static and spiking accuracy; in the reported CIFAR-10 experiments, noise with standard deviation 10 performed better than 20.
  • Smoothing the neuron response makes gradient-based optimization practical while preserving enough of the LIF model’s thresholded behavior for conversion to hard spiking dynamics.
  • Accurate image classification is possible with LIF networks, although the approach relies on offline training, architectural simplifications, and relatively high firing rates compared with biological visual systems.

Summary by alphaXiv.

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