MNIST Neural Network from Scratch

This project implements a fully connected neural network from scratch in Python using NumPy, trained to classify handwritten digits from the MNIST dataset.

---

Features

• Data preprocessing: Normalizes pixel values and one-hot encodes labels.
• Custom neural network framework: Implements Linear layers, activation functions (ReLU, Sigmoid, SoftMax), and loss functions (MSE, NLL).
• Forward and backward propagation: Manual computation of gradients and weight updates.
• Training with SGD: Stochastic Gradient Descent with mini-batch support.
• Evaluation: Computes accuracy on the development set during training.
• Modular design: All layers and losses are implemented as reusable modules in a Sequential container.
• Flexible loss selection: Choose MSE for regression or NLL (cross-entropy) for classification tasks.