Attention-Enhanced CNNs with Contrastive Counterfactual Explanations on MNIST

This repository contains the implementation and research for a project that investigates the integration of attention mechanisms and contrastive counterfactual explanations into convolutional neural networks (CNNs) for digit classification on the MNIST dataset. The goal is to bridge the gap between predictive performance and interpretability using state-of-the-art XAI techniques.

"Understanding why models make mistakes is as important as how accurate they are, and for that we have to see what they 'see' and don't 'see'!"
— Project Motivation

---

Project Overview

In critical domains like healthcare, finance, and law, understanding the reasoning behind AI decisions is crucial. This project aims to make deep learning models more transparent and interpretable without sacrificing accuracy.

Key components:

• Benchmarking and fine-tuning of pretrained CNN backbones (ResNet50, EfficientNetB0, MobileNetV2, etc).
• Integration of attention modules (SE, CBAM, PAM, GC) to enhance interpretability.
• Generation of contrastive counterfactuals to visualize decision boundaries.
• Use of popular XAI techniques including Grad-CAM, Score-CAM, SmoothGrad, and LIME.

---

Experimental Highlights

Module	Accuracy ↑	Test Loss ↓	Params	Notes
ResNet50 + CBAM	98.9%	0.060	~24M	High accuracy, large size
EfficientNetB0+CBAM	97.05%	0.088	~4.4M	Excellent accuracy-to-size ratio
MobileNetV2 + CBAM	97.60%	0.068	~2.6M	Lightweight and fast

Contrastive explanations help answer:

"What minimal change would flip the model’s wrong prediction?"

---

Repository Structure

├── final_notebook.ipynb              # Final implementation and results
├── Attention-Enhanced CNNs...pdf     # Research report and documentation
├── data/                             # Data, metrics CSVs
├── init_research/                    # Initial experiments, baselines, logs
├── notebooks/                        # Research notebooks, ablation studies
├── requirements.txt                  # Libraries and imports used
└── README.md                         # Project overview and usage

---

Methodology

🔍 Model Architectures

• Experiemneting on backbones like VGG, ResNet, EfficientNet, MobileNet on 255x255 RGB-MNIST.
• Attention Wrappers: SE, CBAM, PAM, GC — modularly attached to CNN backbones.
• Fine-tunning the attention-wrapped Final models and exporting them as Native Keras models.

📊 Explainability Techniques

• Attribution Methods: Grad-CAM, Score-CAM, LIME, SmoothGrad
• Counterfactuals:
  • Growing Spheres (black-box)
  • Gradient-based optimization (white-box)
  • Overlay visualizations for human-centered interpretability

🔁 Training & Evaluation

• Optimizer: Adam, LR scheduler: ReduceLROnPlateau
• Data Augmentation: rotation, translation, elastic distortion
• Evaluation Metrics: Accuracy, Test Loss, Perturbation Norm, Visual Clarity

---

Sample Outputs

Original Image	Grad-CAM	Counterfactual	Overlay

Counterfactuals show how close the image was to being classified differently, and the pixels required to flip it's decision.

---

Setup Instructions

1. Clone the repository:

   git clone https://github.com/your_username/xai-contrastive-cnn.git
   cd xai-contrastive-cnn

2. Create a virtual environment:

   python3 -m venv venv
   source venv/bin/activate
   pip install -r requirements.txt

3. Launch Jupyter Notebook and run:

   jupyter notebook notebooks/

Recommended environment: Google Colab with T4 GPU Colab Notebook: xai-counterfactuals

---

The Team

const Contributors = [
{
    name: "🐐 Arindal Char",
    roles: ["Main Analyst", "Main Researcher","Documenter"] },
{
    name: "😊 Trisha Roy Choudhury",
    roles: ["Analyst"] },
{
    name: "📒 Sagnik Hore",
    roles: ["Main Documenter", "Researcher"] },
{
    name: "⌛ Subham Ghosh",
    roles: ["Analyst"] },
{
    name: "🎲 Soumyadeep Biswas",
    roles: ["Analyst", "Documenter"] }
];

---

Contact

For questions, feedback, or collaboration, reach out at:

📧 Mail
🔗 LinkedIn
🐤 X

---

Keep Coding 🚀