Python_MNIST

Overview

After building a simple neural network that recognizes handwritten digits (0-9) from scratch using GitHub Codespaces, Claude Opus 4.5, and the MNIST dataset, this repo was given to Claude Code with a mandate to improve the model.

Result: Model accuracy improved from ~80% to 98.5% on MNIST test set through basic modern deep learning techniques.

Model Architecture

• Input: 784 pixels (28x28 grayscale image)
• Hidden layers: 4 layers with 512 → 384 → 256 → 128 neurons
• Output: 10 neurons (digits 0-9) with softmax
• Activation: ReLU with 30% dropout
• Initialization: He initialization for improved gradient flow

Training Features

• Mini-batch SGD (batch size 128) with data shuffling
• Learning rate scheduling with decay at 50%, 70%, and 90% of training
• 80 epochs with validation monitoring
• Dropout regularization to prevent overfitting

Structure

• network.py - Deep neural network class with forward/backward pass
• train.py - Data loading, training loop with validation
• predict.py - Image preprocessing and prediction with test-time augmentation

Dataset

MNIST - 70,000 labeled images of handwritten digits (28x28 grayscale)

Requirements

• Python 3.x
• NumPy
• Pillow
• SciPy