Florian-BACHO / bats

Error Backpropagation Through Spikes Algorithm

Error Backpropagation Through Spikes (BATS)

Error Backpropagation Through Spikes (BATS) [1] is a GPU-compatible algorithm that extends Fast & Deep [2], a method to performs exact gradient descent in Deep Spiking Neural Networks (SNNs). In contrast with Fast & Deep, BATS allows error backpropagation with multiple spikes per neuron, leading to increased performances. The proposed algorithm backpropagates the errors through post-synaptic spikes with linear time complexity O(N) making the error backpropagation process fast for multi-spike SNNs.
This repository contains the full Cuda implementations of our efficient event-based SNN simulator and the BATS algorithm. All the experiments on the convergence of single and multi-spike models, on the MNIST dataset, its extended version EMNIST and Fashion MNIST are also provided to reproduce our results.

Dependencies and Libraries

Recommanded Python version: >= 3.8

Libraries:

• Cuda (we suggest Cuda 10.1 as this is the version that we used to develop BATS but other versions should also work)

Python packages:

• CuPy [3] (corresponding to the installed version of Cuda)
• matplotlib (Optional. Install only if generate plots with monitors)
• requests (Optional. Install only if run the scripts to download the experiments' datasets)
• scipy (Optional. Install only if run the EMNIST experiment)
• brian2 (Optional. Install only if run the unit tests)

Experiments

Three experiments are available: a single-spike vs multi-spike convergence experiment as well as trainings on the MNIST and EMNIST datasets.

Convergence experiment

$ cd experiments/convergence
$ ls
convergence_count.py  convergence_ttfs.py
$ python3 convergence_count.py
...
$ python3 convergence_ttfs.py
...
$ ls *.pdf
convergence_count.pdf  convergence_ttfs.pdf

Running these python scripts should take a few minutes. The models are trained 100 times for different initial weight distributions. After execution, each script generates a corresponding .pdf file, i.e. convergence_count.pdf and convergence_ttfs.pdf

Download datasets

MNIST

$ cd datasets
$ ls
download_file.py  get_emnist.py  get_fashion_mnist.py  get_mnist.py
$ python3 get_mnist.py
Downloading MNIST...
[██████████████████████████████████████████████████]
Done.
$ ls
download_file.py  get_emnist.py  get_fashion_mnist.py get_mnist.py  mnist.npz

EMNIST

$ cd datasets
$ ls
download_file.py  get_emnist.py  get_fashion_mnist.py  get_mnist.py
$ python3 get_emnist.py
Downloading EMNIST...
[██████████████████████████████████████████████████]
Done.
Extracting EMNIST...
Done.
Cleaning...
Done.
$ ls
download_file.py  emnist-balanced.mat  get_emnist.py  get_fashion_mnist.py  get_mnist.py

Downloading the EMNIST dataset may take a few minutes due to the size of the file.

Fashion MNIST

$ cd datasets
$ ls
download_file.py     get_emnist.py        get_fashion_mnist.py get_mnist.py
$ python3 get_fashion_mnist.py
Downloading Fashion MNIST...
[██████████████████████████████████████████████████]
[██████████████████████████████████████████████████]
[██████████████████████████████████████████████████]
[██████████████████████████████████████████████████]
Done.
$ ls
download_file.py           get_fashion_mnist.py       t10k-images-idx3-ubyte.gz  train-images-idx3-ubyte.gz
get_emnist.py              get_mnist.py               t10k-labels-idx1-ubyte.gz  train-labels-idx1-ubyte.gz

Train models

MNIST

Tree models are available to train with the MNIST dataset:

• a single-spike model (train_ttfs.py)

$ cd experiments/mnist
$ python3 train_ttfs.py
...

• a multi-spike model (train_spike_count.py)

$ cd experiments/mnist
$ python3 train_spike_count.py
...

• and a Convolutional SNN (train_conv.py)

$ cd experiments/mnist
$ python3 train_conv.py
...

During training, plots and data are saved in the output_metrics directory and weights of the best model are saved in the best_model directory.

EMNIST

$ cd experiments/emnist
$ python3 train.py
...

Similarly to the MNIST training, plots are saved in the output_metrics directory and weights of the best model are saved in the best_model directory.

Unit tests

Unit tests can be run by executing the following command:

$ pipenv run python3 -m unittest discover -s tests/ -p "*.py"
.....................
----------------------------------------------------------------------
Ran 21 tests in 74.382s

OK

Some tests use the brian2 [4] clock-based simulator as truth for our event-based simulator.

References

[1] Bacho, F., & Chu, D.. (2022). Exact Error Backpropagation Through Spikes for Precise Training of Spiking Neural Networks. https://arxiv.org/abs/2212.09500
[2] J. Göltz, L. Kriener, A. Baumbach, S. Billaudelle, O. Breitwieser, B. Cramer, D. Dold, A. F. Kungl, W. Senn, J. Schemmel, K. Meier, & M. A. Petrovici (2021). Fast and energy-efficient neuromorphic deep learning with first-spike times. Nature Machine Intelligence, 3(9), 823–835.
[3] Okuta, R., Unno, Y., Nishino, D., Hido, S., & Loomis, C. (2017). CuPy: A NumPy-Compatible Library for NVIDIA GPU Calculations. In Proceedings of Workshop on Machine Learning Systems (LearningSys) in The Thirty-first Annual Conference on Neural Information Processing Systems (NIPS).
[4] Stimberg, M., Brette, R., & Goodman, D. (2019). Brian 2, an intuitive and efficient neural simulator. eLife, 8, e47314.