This repository provides an implementation of Spiking Neural Networks (SNNs) using the SpykeTorch library. SNNs are a type of artificial neural network that more closely mimic the neural mechanisms of the brain, offering potential advantages in computational efficiency and power consumption. This repository includes various SNN architectures and training methods, including Spike-Timing Dependent Plasticity (STDP) and Reward-Modulated STDP (R-STDP).
A Spiking Neural Network (SNN) is an artificial neural network that models neurons more realistically by using spikes to process and transmit information. Unlike traditional artificial neural networks that use continuous values for neuron activations, SNNs operate with discrete events (spikes) that occur at specific points in time.
Spike-Timing Dependent Plasticity (STDP) is a biological learning rule observed in the brain. It adjusts the synaptic strength between neurons based on the relative timing of their spikes. If a presynaptic neuron's spike precedes a postsynaptic neuron's spike, the connection is strengthened (Long-Term Potentiation, LTP). Conversely, if the postsynaptic spike precedes the presynaptic spike, the connection is weakened (Long-Term Depression, LTD).
Reward-Modulated STDP (R-STDP) extends the STDP learning rule by incorporating a reward signal. This reward signal modulates the synaptic changes induced by STDP, allowing the network to learn from both the timing of spikes and an external reward signal, which is useful for reinforcement learning tasks.
This codebase utilizes the SpykeTorch library, which provides tools for implementing and training SNNs with STDP and R-STDP learning rules.
s1c1.py: Contains a class that extracts edges using Differences of Gaussians (DoG) and performs intensity-to-latency encoding.utils.py: Contains utility functions used throughout the project.train.py: The main script for training and evaluating the models.network_trainer.py: A general class that contains methods to train and evaluate SNNs. It serves as a parent class for the models.models/:mozafari2018.py: Contains the SNN architecture explained in the paper by Mozafari et al., 2018.deep2024.py: Implements a deep network with three STDP layers and one R-STDP decision-making layer.deepr2024.py: Contains two networks with two STDP and two R-STDP layers. One network trains all R-STDP layers together, while the other trains them in a greedy manner.inception2024.py: A deep network with an inception-like block that extracts features over different scales.majority2024.py: Similar toinception2024.pybut uses a form of population coding with a majority rule instead of concatenating parallel layers.
- Python 3.11.7
- PyTorch
- SpykeTorch
- Other dependencies as listed in
requirements.txt
Clone the repository:
git clone https://github.com/aidinattar/snn.git
cd snn-training-spyketorchInstall the required dependencies:
pip install -r requirements.txtTo train and evaluate a model, use the train.py script. For example, to train the DeepSNN model on the MNIST dataset:
python train.py --model deep2024 --epochs 2 4 4 10 --tensorboardThis model implements the architecture described in the paper by Mozafari et al., 2018. It uses STDP for training.
A deep SNN with three STDP layers and one R-STDP decision-making layer.
Contains two networks with two STDP and two R-STDP layers. One network trains all R-STDP layers together, while the other trains them in a greedy manner.
A deep network with an inception-like block that extracts features over different scales.
Similar to inception2024.py but uses a form of population coding with a majority rule instead of concatenating parallel layers.
The training history, including accuracy and loss for each epoch, is saved and can be visualized using TensorBoard.
Activation maps for each layer can be optionally saved for further analysis. Note that saving activation maps can take a significant amount of time, and can be disabled if not needed.
The evaluation metrics, including confusion matrix, F1 score, ROC AUC score, and accuracy, are computed and saved during the evaluation phase.
This repository is based on the work of Mozafari et al. and utilizes the SpykeTorch library for SNN training and evaluation.
This project is licensed under the MIT License - see the LICENSE file for details.