🧠 Spiking Neural Networks (SNN) for Classification using snnTorch

This project demonstrates the implementation and training of a Spiking Neural Network (SNN) on the MNIST dataset using the snntorch library. Inspired by the behavior of biological neurons, SNNs are energy-efficient and time-aware neural networks, ideal for sequential data processing.

---

Overview

• ANN (Artificial Neural Network): Uses continuous (numeric) signals.
• SNN (Spiking Neural Network): Uses discrete spikes that mimic real neuron firing.
• Neuron Model Used: Leaky Integrate-and-Fire (LIF)
• Dataset: MNIST Handwritten Digits
• Library: snntorch

---

Project Structure

1. Data Loading
   Loads the MNIST dataset using torchvision.datasets and converts it to tensors.

2. Model Definition
   A fully connected 2-layer SNN with LIF neurons:

   fc1 = nn.Linear(784, 512)
   lif1 = snn.Leaky(beta=0.5)
   fc2 = nn.Linear(512, 10)
   lif2 = snn.Leaky(beta=0.5)

3. Training

   • Inputs are propagated through the network for a fixed number of time steps.
   • Spike outputs are averaged over time.
   • Uses CrossEntropyLoss and Adam optimizer.
   • Surrogate gradients are applied to enable training.

4. Evaluation

   • Final predictions are obtained by averaging spikes across time.
   • Model accuracy is calculated on the test set.

5. Confusion Matrix

   • A normalized confusion matrix is visualized using seaborn.

📈 Results

• Accuracy: ~96-98% on MNIST after a few epochs
• Confusion Matrix: Visualizes performance across digit classes