BMVC2022_SVPG

This repository contains some basic (easy to run) codes for the experiments in the BMVC 2022 paper:

Biologically Plausible Variational Policy Gradient with Spiking Recurrent Winner-Take-All Networks

Citation at the bottom of this link to paper.

Note: An updated version of the experiments is available at this repository.

Prerequisites

System requirements: tested on Ubuntu 18.04, NVIDIA Geforce 3080, 32GB RAM, Anaconda3.

Part of installed packages: python(3.6), torch(1.8.2), snntorch(0.4), torchvision(0.9), scikit-image(0.17), opencv-python(4.5)

General running:

Simply run python <code name>.py to run the experiment with default parameters.

You may need to download the MNIST dataset. A preprocessing code is provided in BMVC2022_SVPG/MNIST/MNIST_DATA/. Here are some of the parameters. See parser in the codes for more settings.

• --cuda sets the GPU to use;

• --rep sets the random seed;

Logs are stored in the ./log/ folder. Create it if it does not exist.

Explanations

MNIST

MN_SVPG.py, MN_SVPG_shrink.py:

Python codes respectively for the SVPG and SVPG-shrink methods.

MNWTArate_nop.py, MNWTAuni_nop.py, MNWTAdexp_nop.py:

Python codes for the comparison of the three implementations of SVPG, respectively rate coded with noise, spike coded with rectangle window, and spike coded with double exponential window.