surrogate-gradient

The comprehensive bilingual (EN/中文) hub for Spiking Neural Networks — 340+ papers, models, neuromorphic hardware, datasets, tools & research groups.

Minimal PyTorch examples for the four-stage structural evolution from ANN to event-driven SNN: Stage 0 (baseline ANN) → Stage 1 (binarization) → Stage 2 (temporal expansion) → Stage 3 (temporal accumulation) → Stage 4 (reset & sparsity control).

A spiking model of macaque MT and LIP that learns perceptual decisions with surrogate gradients and dopamine-modulated STDP, with an interactive 3D brain.

Official code for "Attacking the Spike: On the Security of Spiking Neural Networks to Adversarial Examples" (Neurocomputing 2025). Implements the MDSE adversarial attack for SNNs, CNNs, and Vision Transformers.

Brian2-compatible spiking neural network simulation and native training with a Rust engine, CPU/GPU/browser/MPI backends, AtlasIR, and experimental neuromorphic model conversion.

The whole male Drosophila CNS connectome (166,700 neurons) used uncut as the spiking core of a language model. Scope, numbers, honest controls, pipeline animation.

FS²-SSA: a ~94M causal spiking language model on FineWeb-Edu. Softmax-free attention, every matrix product of the blocks fed by few-spike neurons, and an exact linear recurrence for inference.

EpiPredict by Samuelson G: Multimodal deep learning (CNN-LSTM, Transformer) & neuromorphic SNN edge framework for epileptic seizure prediction using wearable EEG, fNIRS, autonomic sensors & big data. Includes paper & interactive web console.

Ava: a brain-like language model in pure Rust. Spiking neurons, fast synapses, sparse codes, and sleep with synaptic consolidation. No position embeddings, almost no attention. Tiny today, built to scale. Source-available (PolyForm Noncommercial).

Device-to-algorithm neuromorphic: train spiking neural nets in snnTorch, then deploy on a simulated memristor crossbar with measured SnS2 device non-idealities.

Reproducibility code and revision artifacts for neuron heterogeneity, input representation, and dynamical stability in spiking networks.

Store, search, and reason over data in spike timing — spiking neural networks for data storage & retrieval. Less compute, less storage. Three paradigms: telemetry hub, in-storage query engine, relational spike-time embeddings.