NeuroCore is reusable building blocks and training idioms for Spiking Neural Networks (SNNs) in PyTorch. The repository focuses on vibration classification: HUMS-style environments and CWRU bearing data.
Stateful temporal modeling
A standard feed-forward net is stateless. An SNN is not. Temporal state has to be kept across time steps: membrane potentials and spike trains. Stateful blocks are first-class primitives here. Training loops and dataset loaders handle sequences where the signal is energy evolving over windows, not a static feature vector.
Explicit surrogate gradients
Binary spikes are non-differentiable. Backprop needs a smooth approximation. The codebase makes surrogate gradients an explicit, tunable parameter, not a hidden layer. Surrogate choice and stiffness change learning stability. The repo gives one experimental surface to compare those settings without rewriting training boilerplate.
Honest baselines
The project enforces rigorous benchmarking by keeping classical feature-based classifiers in the same repository as the SNN models. Improvements in the spiking architecture are measured against those baselines, so the question is whether the SNN is actually better, not only whether it is novel.