spikingtorch

Training spiking neural networks using Pytorch

About

spikingtorch is a lightweight package for training deep neural networks using Pytorch. spikingtorch includes encoders that transform standard ML datasets into spike trains, and decoders that transform the output spikes into values that can be used with loss functions in Pytorch.

spikingtorch implements spiking neural networks and backpropagation through spikes for leaky integrate and fire neurons.

In addition to the Python package, this repository also implements the methodology and reproduces the results presented in the paper:

A. Yanguas-Gil, Coarse scale representation of spiking neural networks: backpropagation through spikes and application to neuromorphic hardware, arXiv:2007.06176

Status

spikingtorch is in active development, with more neuron models coming up soon.

Quick install

Through pypi:

pip install spikingtorch

Acknowledgements

• Threadwork, U.S. Department of Energy Office of Science, Microelectronics Program.

The original implementation was based on reseach funded through Argonne's Laboratory Directed Research and Development program.

Publications

Spikingtorch's backpropagation approach is based on the following work:

A. Yanguas-Gil, Coarse scale representation of spiking neural networks: backpropagation through spikes and application to neuromorphic hardware, arXiv:2007.06176

Copyright and license

Copyright © 2023, UChicago Argonne, LLC

spikingtorch is distributed under the terms of BSD License. See LICENSE