Digit Recognition using Neural Networks

A Python implementation of a neural network for recognizing handwritten digits using the MNIST dataset.

Overview

This project implements a feedforward neural network from scratch to classify handwritten digits. The model achieves approximately 98% accuracy on the MNIST test set.

Features

• Custom implementation of neural network architecture
• Support for different cost functions (MSE and Cross Entropy)
• Stochastic gradient descent optimization
• Early stopping to prevent overfitting
• Batch processing for efficient training

Dependencies

• Python 3.8+
• NumPy
• tqdm
• pickle
• gzip

Dataset

The project uses the MNIST dataset, which should be downloaded as mnist.pkl.gz and placed in the project root directory.

Usage

1. Ensure all dependencies are installed
2. Download the MNIST dataset
3. Run the Jupyter notebook Digit_Recognition.ipynb

Model Architecture

• Input layer: 784 neurons (28x28 pixel images)
• Hidden layer 1: 300 neurons
• Hidden layer 2: 100 neurons
• Output layer: 10 neurons (one for each digit)

Performance

The model typically achieves:

• Training accuracy: ~98%
• Test accuracy: ~98%