SpikeNN

Spiking Neural Network (SNN) framework for classification, featuring fully-connected architectures, first-spike coding, single-spike IF/LIF neurons, floating-point spike timestamps (event-driven), and STDP-based supervised learning rules.

This is the code for the paper Neuronal Competition Groups with Supervised STDP for Spike-Based Classification, published at NeurIPS 2024 (see official repo).

Getting started

The code is written with Python3 (3.8.10) and runs exclusively on CPUs.

Requirements

• numpy (1.23.4)
• numba (0.56.4)
• tqdm (4.64.1)
• setuptools (45.2.0)

Install

The spikeNN package and its dependencies can be installed with:

python3 -m pip install -e .

In case you get any errors with a dependency, please try to use the same version as ours.

Usage

Applications are located in the app/ folder.
Configuration files are located in the config/ folder.
The spikeNN package provides all the core classes and functions (see spikenn/).
The Readout class (app/readout.py) defines a flexible SNN model for classification using the spikeNN package.

Input data

This framework is designed to process input data wrapped into SpikingDataset objects, where each sample is represented by an ordered sparse list of spikes. A spike is a tuple <i, t> (i: 1D-index of the input neuron, t: firing timestamp).

The easiest way to convert a dense numpy dataset (here, training set) into a SpikingDataset is as follows:

python3 app/generate_dataset.py /input/X_train.npy /input/y_train.npy /output/trainset.npy 

• X_train.npy: numpy array of shape (N_samples, D1, ..., DN), where each value is a spike timestamp.
• y_train.npy: numpy array of shape (N_samples,), where each value is a class label.

Run

To start a single run:

python3 app/run.py /input/data/dir/ /output/data/dir/ /config/file [--seed 0]

To start a K-fold run:

python3 app/kfold.py /input/data/dir/ /output/data/dir/ /config/file [--K 10] [--n_proc 10]

When the run is done, you can use python3 app/get_kfold_score.py /output/data/dir/ to compute the mean accuracy.

To start a gridsearch run:

python3 app/gridsearch.py /input/data/dir/ /output/data/dir/ /config/file [--n_proc MAX] [--resume]

When the run is done, you can use python3 app/get_gs_best.py /output/data/dir/ to show the best gridsearch runs.

Notes:

• Input directory must contain SpikingDataset files with the following name convention: trainset.npy, valset.npy (optional), testset.npy (optional).
• Configuration files use the JSON format and follow a specific syntax (see examples in config/).
• Use python3 <app_file>.py -h for additional information regarding arguments.

Reproduce the main results

All our quantitative results are computed with K-fold runs.
All the configuration files employed in our experiments are located in the config/ folder.
Gridsearch-optimized configurations are placed in the config/<FT-EXTRACTOR>/<DATASET>/ folders and gridsearch configurations in the gs/ sub-folders.

Unsupervised feature extraction

In the paper, datasets are preprocessed with Hebbian-based unsupervised feature extractors before classification:

• STDP-CSNN: a single-layer spiking CNN trained with STDP (original repository).
• SoftHebb-CNN: a three-layer non-spiking CNN trained with SoftHebb (original repository).

The folder ft-extract/ provides the scripts and documentation for obtaining and preprocessing each dataset with these feature extractors, as well as converting them to SpikingDataset objects.

We also provide some features extracted with STDP-CSNN to facilitate quick experimentation: MNIST, Fashion-MNIST, CIFAR-10.

5.2 Accuracy comparison

The following command can be used to reproduce our scores from Table 1:

python3 app/kfold.py ft-extract/extracted/<FT-EXTRACTOR>/<DATASET>/kfold/ logs/<FT-EXTRACTOR>/<DATASET>/kfold/<CONFIG>/ config/<FT-EXTRACTOR>/<DATASET>/<CONFIG>.json

• Replace <DATASET> by: MNIST, F-MNIST, CIFAR10, CIFAR100.
• Replace <FT-EXTRACTOR> by: STDP-CSNN, SoftHebb-CNN.
• Replace <CONFIG> by: s2stdp+ncg, sstdp+ncg, s2stdp, sstdp, rstdp-5n, rstdp-20n.

Output will be logged in .txt files at logs/<FT-EXTRACTOR>/<DATASET>/kfold/<CONFIG>/.

5.3 Ablation study

Here are the names of the configuration files used in for the ablation study on S2-STDP+NCG:

• M-1: s2stdp.json
• M-5: s2stdp+ncg-no-cr-no-nt.json
• M-5+CR-1: s2stdp+ncg-cr-1-thr-no-nt.json
• M-5+CR-2: s2stdp+ncg-no-nt.json
• M-5+CR-1+L: s2stdp+ncg-cr-1-thr.json
• M-5+CR-2+L: s2stdp+ncg.json
• M-5+Drop+L: s2stdp+ncg-cr-drop.json

5.4 Impact of competition regulation

In app/readout.py, the variable self.save_stats must be set to True to save the training logs used to generate the figures.

Limitations

• Limited to first-spike coding (TTFS) with single-spike IF/LIF neurons
• No support for other architectures than fully-connected
• Multi-layer training not implemented
• Long computation times (CPU-based)

Known bugs

• Results may vary slightly, though not significantly, depending on the version of Numba installed.

Note: if you encounter any issues with the code, please report them by creating an issue.

Acknowledgments

• Chaire Luxant-ANVI (Métropole de Lille)
• UAR 3380 - IRCICA - Institut de recherche sur les composants logiciels et matériels pour l'information et la communication avancée, Lille, F-59000, France
• Univ. Lille, CNRS, Centrale Lille, MR 9189 - CRIStAL - Centre de Recherche en Informatique, Signal et Automatique de Lille, Lille, F-59000, France