jeshraghian/snntorch

Deep and online learning with spiking neural networks in Python

What it solves

Modern deep learning typically uses continuous values to process information, which is computationally expensive. The human brain, however, is highly efficient because it encodes information in discrete "spikes." snnTorch bridges this gap by allowing developers to build and train Spiking Neural Networks (SNNs) using the familiar PyTorch ecosystem.

How it works

snnTorch integrates spiking neuron models directly into PyTorch as recurrent activation units. Because these models are represented by recursive functions, the system does not need to store membrane potential traces for every neuron to calculate gradients, making it memory-efficient. It leverages PyTorch's native GPU acceleration for fast computation and supports surrogate gradient descent to enable training via backpropagation.

Who it’s for

This tool is designed for researchers and engineers who want to explore neuromorphic computing or build energy-efficient neural networks using PyTorch. It is suitable for those familiar with standard deep learning workflows who wish to transition to spike-based architectures.

Highlights

  • Deep integration with PyTorch, allowing spiking neurons to be used like standard layers (e.g., convolutional or fully-connected).
  • Built-in support for spike generation, data conversion, and visualization.
  • Interoperability with other SNN libraries via the Neural Intermediate Representation (NIR) format.
  • Includes specialized tools for spiking-aware weight initialization and surrogate gradient functions.
  • Extensive library of tutorials and notebooks covering everything from basic neuron models to advanced regression and classification tasks.

Related

  • Project
    deepmodeling/deepmd-kit

    A deep learning package that turns quantum-mechanical reference data into fast, scalable interatomic potentials for molecular and materials science simulations.

  • Project
    chenglou/handwired

    Handwired converts regular JavaScript code into a differentiable program made only of matrix multiplication and ReLU, enabling GPU‑accelerated execution and gradient‑based experiments without any training.

  • Project
    WenjieDu/PyPOTS

    PyPOTS is a Python machine learning toolbox designed to handle partially-observed multivariate time series data, providing unified APIs for imputation, forecasting, classification, clustering, and anomaly detection.

  • Project
    gavinkhung/machine-learning-visualized

    Machine Learning Visualized is a collection of interactive Jupyter Notebooks and documentation that explains machine learning algorithms through mathematical derivations and visual training simulations.