Spiking Neural Networks with PyTorch & Norse

Anastasiya Kotelnikova

Completed work

Researcher

Matplotlib

PyTorch

seaborn

Education

Spiking Neural Networks with PyTorch & Norse

This repository contains the implementation of Spiking Neural Networks (SNNs) using PyTorch and the Norse library. This project was researched, proposed, and led by Anastasiya Kotelnikova and Mehrvish Mirza as part of the DS677 Deep Learning course at NJIT. It explores neuromorphic computing by implementing biologically inspired models and comparing SNNs to traditional Artificial Neural Networks (ANNs).

Project Overview

Spiking Neural Networks (SNNs) are a biologically inspired class of neural networks that simulate neuron behavior using discrete spikes for energy-efficient and event-driven processing.

This project focuses on building and evaluating SNNs using time-based datasets and benchmarking them against standard deep learning models.

Objectives

Implement an SNN model using PyTorch and Norse

Preprocess and use neuromorphic dataset: SHD

Compare performance of SNNs vs. traditional ANNs

Optimize SNN hyperparameters for improved training accuracy and efficiency

Document results and provide research insights into neuromorphic computing

Repository Structure

data/            
models/          
notebooks/       # Colab notebooks for experimentation and prototyping
reports/         # Milestones, Reports, Saved checkpoints and trained models, Datasets used for training and testing
src/             
results/

Tools & Libraries

Python 3.10

PyTorch

Norse (github.com/norse/norse)

NumPy, Matplotlib, Seaborn

Google Colab (for experimentation)

Sample Results

Dataset Model Accuracy Notes SHD SNN 84% LIFCell-based SNN N-MNIST SNN 92% Spike-trained SNN MNIST ANN 98% CNN baseline

Additional outputs and performance logs are available in the notebooks/ folder.

Key Concepts

Neuromorphic Computing: Simulating brain-like neural behavior using spikes

LIF Neurons: Biologically inspired neurons with leak-integrate-fire dynamics

SNN vs ANN: Compared accuracy, latency, and training performance

Spike Encoding: Converting continuous input into time-based spikes

Backpropagation Through Time (BPTT): Used to train spike-based models

Contributors

Anastasiya Kotelnikova

Master’s Student in Data Science at NJIT anastasiya.kotelnikova21@gmail.com

Mehrvish Mirza

AI Certificate Student at NJIT | Analyst with a background in Finance & IT Currently enrolled in DS677 – Deep Learning (Spring 2025) alongside the author