Handwritten Digit Classification with PyTorch

This project presents handwritten digit classification on the MNIST dataset using several neural network architectures implemented in PyTorch.

The notebook compares simple and more advanced models, analyzes overfitting, applies regularization, and tunes a convolutional neural network to achieve more than 99% test accuracy.

Models included

• Perceptron — a single-layer linear baseline
• Deep fully-connected network — a model with a hidden layer
• CNN — a basic convolutional neural network
• BetterCNN — a tuned convolutional model with Batch Normalization, dropout, and Adam optimizer

Project goals

• load and preprocess the MNIST dataset
• split the data into training, validation, and test subsets
• train and compare multiple neural network models
• analyze overfitting using training and validation loss
• apply regularization techniques such as L2 regularization
• tune a CNN model to exceed 99% test accuracy
• visualize classification errors using a confusion matrix

Dataset

The project uses the MNIST dataset of handwritten digits:

• 60,000 training images
• 10,000 test images
• grayscale images of size 28 × 28
• 10 classes representing digits from 0 to 9

In this project, the original training set is further split into:

• 50,000 training samples
• 10,000 validation samples

Results summary

Model	Test Accuracy	Notes
Perceptron	~92%	Simple baseline, limited capacity
Deep	~96%	Better performance, but shows overfitting
Deep + L2	~96%	Slightly more stable validation behavior
CNN	~96–98%	Better suited for image classification
BetterCNN	99.27%	Best model, exceeded the target

Key observations

• The Perceptron works as a useful baseline, but it is too simple for best performance.
• The Deep model significantly improves the results, but it tends to overfit.
• Adding L2 regularization slightly improves generalization.
• CNN-based models perform better because they preserve the spatial structure of image data.
• The tuned BetterCNN achieved 99.27% test accuracy, successfully exceeding the required 99% threshold.
• The confusion matrix shows that the remaining mistakes mostly occur between visually similar digits such as 9 and 5, 9 and 4, or 7 and 2.

Repository structure

mnist-classification-pytorch/
├── README.md
├── requirements.txt
├── .gitignore
├── notebooks/
│   └── mnist_classification.ipynb
└── images/