MNIST Classification with Neural Networks

This repository contains two implementations of neural networks for handwritten digit classification on the MNIST dataset using PyTorch.
The project compares a fully connected baseline (MLP) with a convolutional neural network (CNN) to highlight the advantages of convolutional architectures for image tasks.

---

Results

• MLP (Multi-Layer Perceptron): ~97% test accuracy.
• CNN (Convolutional Neural Network): ~99% test accuracy.
• Training and validation curves show significant generalization improvement with CNNs.

---

Features

• Data preprocessing and normalization using torchvision.
• Two separate implementations: MLP (dense layers) and CNN (convolutions + pooling).
• Training pipeline with Adam optimizer and cross-entropy loss.
• Evaluation using accuracy and confusion matrices.
• Modular Jupyter notebooks for interactive exploration.

---

Repository Structure