Handwritten Digit Recognition from Scratch

A neural network built from scratch using pure Python to recognize MNIST handwritten digits with 95-97% accuracy — no NumPy, TensorFlow, or PyTorch!

🎯 Features

• Pure Python implementation (no ML libraries)
• Custom matrix operations from scratch
• Forward & backward propagation
• Mini-batch gradient descent
• 95-97% test accuracy on MNIST

🏗️ Architecture

Input (784) → Hidden (128, ReLU) → Output (10, Softmax)

🚀 Quick Start

# Clone repository
git clone https://github.com/kaushalrog/digitrecognition.git
cd digitrecognition

# Install matplotlib (only dependency)
pip install matplotlib

# Download MNIST data
python download_mnist.py

# Train model
python train.py

📊 Results

• Training Accuracy: ~96-98%
• Test Accuracy: ~95-97%
• Training Time: ~5-10 minutes (20 epochs)

📁 Files

• my_math.py - Custom matrix operations
• data_loader.py - MNIST loader
• neural_network_scratch.py - Neural network
• train.py - Training script

🧠 What's Inside

Built from scratch:

• Matrix class with all operations
• ReLU & Softmax activations
• Cross-entropy loss
• Backpropagation algorithm
• He weight initialization
• Mini-batch gradient descent

👨‍💻 Author

Kaushal - GitHub

📝 License

MIT License

---

⭐ Star if you found this helpful!