https://github.com/jeremiegince/snnimageclassification
Comparison of neural dynamics like LIF and ALIF for image classification on MNIST and Fashion-MNIST datasets.
https://github.com/jeremiegince/snnimageclassification
deep-learning fashion-mnist mnist neuroscience python3 pytorch snn spiking-neural-networks
Last synced: 7 months ago
JSON representation
Comparison of neural dynamics like LIF and ALIF for image classification on MNIST and Fashion-MNIST datasets.
- Host: GitHub
- URL: https://github.com/jeremiegince/snnimageclassification
- Owner: JeremieGince
- License: mit
- Created: 2022-02-09T15:47:20.000Z (over 3 years ago)
- Default Branch: main
- Last Pushed: 2022-05-04T03:35:40.000Z (over 3 years ago)
- Last Synced: 2025-01-20T14:34:56.402Z (9 months ago)
- Topics: deep-learning, fashion-mnist, mnist, neuroscience, python3, pytorch, snn, spiking-neural-networks
- Language: Python
- Homepage:
- Size: 1.65 MB
- Stars: 1
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
Classification of MNIST and Fashion-MNIST images using SNNs
## Abstract
Spiking neural networks (SNN) are a new approach to artificial intelligence, which try to mimic biological
neural networks by using neural dynamics from neuroscience, such as LIF and ALIF. This project has thus
allowed us to highlight the potential and efficiency of SNNs in artificial intelligence and specifically
in image classification. Indeed, the image databases, MNIST and Fashion-MNIST, were classified with this
type of network, with respective accuracies of 96.19% and 81.94%, which is competitive with multilayer
perceptron networks (MLP) (95.3% for MNIST and 91.4% for Fashion-MNIST). Moreover, it was possible to
demonstrate that the ALIF dynamics is more efficient than the LIF dynamics for image classification. Other
parameters also have an important effect on the accuracy of the models: the presence of recurrent connections
decreases the accuracy by an average of 45%, the use of a periodic input signal decreases it by 15%
and the number of neurons in the network has no significant effect on the accuracy. Finally, ALIF was able
to outperform its state-of-the-art MLP counterpart for the MNIST dataset. For Fashion-MNIST, although the
SNN models did not outperform the state-of-the-art results, several avenues of research were revealed that
would significantly improve the results, such as extending the training, training the beta parameters, and
reducing the number of neurons.
## Prediction Pipeline
The prediction pipeline for the spiking neural network is as follows:
in the previous figure, the x variable is the input image, the Sx variable is the image transformed
into spikes, the z variable is the spikes generated by the SNN, and the y variable is the output of the
readout layer. Finally, the classification probability px are computed by the softmax function.
## Results
The results of the experiments are presented in the following figures.
### MNIST
### Fashion-MNIST
### Legend
- REC : recurrent connections flag;
- P : periodical signal flag;
- H or HN : number of hidden neurons;
- I : number of training iterations;
- B : beta training flag.
## Requirements
- ```pip install -r requirements.txt```
- ```pip install git+https://github.com/JeremieGince/PythonBasicTools```
## Code Example
Code example for the classification of the MNIST dataset using an SNN with the ALIF dynamics.
```python
import os
import pprint
from typing import Any, Dict
import psutil
from src.datasets.datasets import DatasetId, get_dataloaders
from src.modules.snn import LoadCheckpointMode, SNN
from src.modules.spike_funcs import SpikeFuncType
from src.modules.spiking_layers import LayerType
from src.modules.training import hash_params, save_params
def train_with_params(params: Dict[str, Any], data_folder="tr_results", verbose=False):
checkpoints_name = str(hash_params(params))
checkpoint_folder = f"{data_folder}/{checkpoints_name}"
os.makedirs(checkpoint_folder, exist_ok=True)
dataloaders = get_dataloaders(
dataset_id=params["dataset_id"],
batch_size=256,
to_spikes_use_periods=params["to_spikes_use_periods"],
nb_workers=psutil.cpu_count(logical=False),
)
network = SNN(
inputs_size=28 * 28,
output_size=10,
n_hidden_neurons=params["n_hidden_neurons"],
spike_func=params["spike_func"],
hidden_layer_type=params["hidden_layer_type"],
use_recurrent_connection=params["use_recurrent_connection"],
checkpoint_folder=checkpoint_folder,
learn_beta=params.get("learn_beta", True),
)
save_params(params, os.path.join(checkpoint_folder, "params.pkl"))
network.fit(
dataloaders["train"],
dataloaders["val"],
nb_epochs=params.get("nb_epochs", 15),
load_checkpoint_mode=LoadCheckpointMode.LAST_EPOCH,
force_overwrite=False,
verbose=verbose,
)
network.load_checkpoint(LoadCheckpointMode.BEST_EPOCH)
return dict(
network=network,
accuracies={k: network.compute_classification_accuracy(dataloaders[k]) for k in dataloaders},
checkpoints_name=checkpoints_name,
)
if __name__ == '__main__':
results = train_with_params(
{
"dataset_id": DatasetId.MNIST,
"to_spikes_use_periods": True,
"n_hidden_neurons": 128,
"spike_func": SpikeFuncType.FastSigmoid,
"hidden_layer_type": LayerType.ALIF,
"use_recurrent_connection": True,
"learn_beta": True,
"nb_epochs": 30,
},
verbose=True,
)
pprint.pprint(results, indent=4)
```
The output of the previous code:
```
|####################################################################################################| 30/30, [00:39:33<00:00:00, 0itr/s] 100% train_loss: 1.88297e-01, val_loss: 2.78507e-01
{ 'accuracies': { 'test': 0.9366999864578247,
'train': 0.9565490484237671,
'val': 0.9359999895095825},
'checkpoints_name': '285720764938101059982133148144568371179',
'network': SNN(
(layers): ModuleDict(
(input): ALIFLayer()
(readout): ReadoutLayer()
)
)}
```
# License
[MIT License](LICENSE.md)
# Citation
```
@article{Gince_LamontagneCaron_SNNImgClassification_2022,
title={SNN Image Classification},
author={Gince, Jérémie and Lamontagne-Caron, Rémi},
year={2022},
publisher={Université Laval},
url={https://github.com/JeremieGince/SNNImageClassification},
}
```