kclip / VDIB

snn

Code for Learning to Time-Decode in Spiking Neural Networks Through the Information Bottleneck.

Run example

Code for our experiments can be run from the scripts

python image_reconstruction.py

python mnistdvs_reconstruction.py

python predictive_coding_softmax.py

Data preprocessing

Preprocessing/loading for the MNIST-DVS dataset requires the neurodata module from github.com/kclip/neurodata

Dependencies

numpy v1.18.1

pytorch v1.7.1

tables v3.6.1

scipy v1.4.1

neurodata (from github.com/kclip/neurodata)

snn (from github.com/kclip/snn) - up to commit 2da2eaa8626c8d361783fb3b9488181087cf6c33

Support

The master branch contains the paper version of the code for reproducibility. This version is now obsolete, please have a look at the eprop-fixed branch for an updated version, including improved performance, support of minibatching, and fixes in the implementation of e-prop.

Written by Nicolas Skatchkovsky