Digit Classification with Explainable AI (XAI)

A PyTorch-based CNN model for handwritten digit classification with integrated Explainability (Grad-CAM).

---

Overview

This project implements a Convolutional Neural Network (CNN) for grayscale digit classification (28×28 images). It follows a clean and modular design suitable for research, teaching, and reproducibility. In addition to classification, it integrates explainable AI techniques (Grad-CAM) to visualize model reasoning and highlight the regions influencing predictions.

---

📁 Project Structure

Digit_Classification_XAI/
│
├── config/
│ └── config.yaml # Hyperparameters and training configuration
│
├── data/
│ ├── train/ # Training images organized by class
│ └── test/ # Test images organized by class
│
├── src/
│ ├── main.py # Entry point for training and evaluation
│ ├── models/
│ │ └── cnn_model.py # CNN model definition
│ ├── utils/
│ │ ├── dataset.py # Data loading and augmentation
│ │ ├── train_eval.py # Training and evaluation utilities
│ │ └── explain.py # Grad-CAM and visualization methods
│
├── notebooks/
│ └── EDA_and_Explainability.ipynb # Exploratory analysis and XAI demonstrations
│
├── checkpoints/ # Saved model weights
├── README.md
└── requirements.txt

---

Model Architecture

---

Evaluation Results

Confusion Matrix

Classification Report (Macro Averages)

Metric	Score
Accuracy	0.9600
Precision	0.9607
Recall	0.9600
F1-Score	0.9598

---

Explainability (Grad-CAM)

Grad-CAM visualizations show where the CNN focuses when making predictions. Misclassified samples are analyzed to understand model bias or confusion.

---

How to Run

1. Clone the repository

git clone https://github.com/rasoulameri/Digit_Classification_XAI.git
cd Digit_Classification_XAI

2. Create environment

conda create -n pytorch_gpu python=3.10
conda activate pytorch_gpu
pip install -r requirements.txt

3. Train the model

python -m src.main

4. Explore explainability in Jupyter

jupyter notebook notebooks/EDA_and_Explainability.ipynb

Key Features:

• Modular PyTorch training pipeline
• Config-driven architecture
• Integrated Grad-CAM explainability
• Deterministic and reproducible results
• Compatible with CPU and GPU

---

📫 Contact

Rasoul Ameri
📧 rasoulameri90@gmail.com
🔗 GitHub Profile

---