A neural network implementation from scratch using NumPy to classify handwritten digits (1, 2, and 3) from the MNIST dataset.

• Project Overview
• Repository Structure
• Dependencies
• Implementation Details
• Results
• Credits

This project implements a simple neural network from scratch (using NumPy) to classify handwritten digits (1, 2, and 3) from the MNIST dataset. The implementation includes:

• Data loading and preprocessing
• One-hot encoding of labels
• Neural network with single layer
• Softmax activation and cross-entropy loss
• Training with gradient descent
• Performance evaluation

mnist-digit-classification/
├── mnist_neural_network.py # Main implementation script
├── README.md # This documentation file
├── requirements.txt # Python dependencies
├── images/ # Directory for output plots
│ ├── accuracy_vs_epochs.png
│ ├── accuracy_vs_lr.png
│ └── loss_vs_epochs.png
└── data/ # MNIST dataset (auto-downloaded)

• Python 3.6+
• NumPy
• Matplotlib
• Keras (for dataset loading)

Install requirements:

pip install -r requirements.txt

Key components of the implementation:

1. Data Preparation

• Load MNIST dataset using Keras
• Filter for only digits 1, 2, and 3
• Flatten and normalize images (28x28 → 784-dim vectors)
• One-hot encode labels

2. Neural Network

• Single layer architecture (784 input → 3 output)
• ReLU activation function
• Softmax output layer
• Cross-entropy loss function

3. Training

• Batch gradient descent
• Learning rate comparison (0.1 to 0.5)
• 50 training epochs

4. Evaluation

• Test accuracy calculation
• Loss and accuracy tracking
• Prediction visualization

The model achieves the following performance:

Learning rate comparison results:

• MNIST dataset from Yann LeCun and Corinna Cortes
• NumPy for numerical computations
• Keras for dataset loading utilities