Bio-Inspired-Computer-Vision (Still in Progress)

This repository contains code for training and evaluating a spiking neural network (SNN) with a spiking RNN region‑of‑interest (ROI) predictor, alongside a CNN baseline, on voxelised event‑based vision data.

Dissertation project

This work forms part of a dissertation at the University of Manchester. It explores how SNNs can leverage ROI prediction to improve classification on event‑based vision datasets.

Features

• Spiking RNN ROI predictor integrated with a LeNet‑style SNN
• CNN baseline for comparison
• Real‑time progress bars with tqdm for clear training and evaluation feedback
• Class‑weighted loss to balance skewed class frequencies
• Evaluation scripts producing confusion matrices and classification reports

Repository structure

.
├── data               # Dataset loaders and transforms
│   ├── datasets.py
│   └── transforms.py
├── models             # Network definitions
│   ├── spiking_rnn_roi.py
│   ├── lenet_snn.py
│   └── cnn_baseline.py
├── scripts            # Training and evaluation scripts
│   ├── train_snn.py
│   ├── train_cnn.py
│   └── evaluate.py
├── results            # Saved models and output figures
│   ├── figures
│   └── reports
└── README.md

Installation

1. Clone this repository:

   git clone https://github.com/yourusername/neuromorphic-classification.git
   cd neuromorphic-classification

2. Create and activate a virtual environment.

3. Install dependencies:

   pip install -r requirements.txt

Data preparation

1. Download the NMNIST or DVS Gesture dataset.
2. Update paths in data/datasets.py to point to your data directory.
3. The code converts events into 5D voxel tensors of shape [batch, time, channels, height, width].

Training

Spiking RNN ROI predictor + SNN

python scripts/train_snn.py \
  --epochs 5 \
  --batch-size 32 \
  --lr 1e-3 \
  --num-workers 2

This script displays batch‑level progress using tqdm and saves the best model under results/.

CNN baseline

python scripts/train_cnn.py \
  --epochs 5 \
  --batch-size 32 \
  --lr 1e-3 \
  --num-workers 2

Evaluation

Run the evaluation script to generate confusion matrices and classification reports:

python scripts/evaluate.py \
  --model-path results/snn_best.pth \
  --dataset test

The script prints a classification report to the console and saves a confusion matrix plot in results/figures.

Results

Training the SNN improved accuracy from around 17% to nearly 68% over five epochs. The confusion matrix highlighted bias towards certain classes, prompting the use of class weights.

Future work

• Explore alternative ROI predictor architectures
• Add event‑data augmentation strategies
• Investigate online learning and adaptive threshold techniques

License

This project is licensed under MIT.