Neural Network from Scratch - Handwritten Digit Recognition

This project implements a fully connected neural network from scratch (using only NumPy) to recognize handwritten digits from the MNIST dataset.
No deep learning frameworks like TensorFlow or PyTorch are used — everything from forward propagation to backpropagation is coded manually.

A simple fully connected feedforward neural network with one hidden layer.

• Manual coding of forward propagation, backpropagation, and gradient descent.
• No use of high-level deep learning libraries—everything is written from scratch in NumPy
• Achieved an accuracy of approximately 85%

Project Overview

• Built a multi-layer neural network (with activation functions, loss function, and gradient descent).
• Trained on the MNIST dataset (28x28 pixel handwritten digits).
• Achieved good accuracy without using high-level ML libraries.
• Visualized training performance and predictions.

How It Works

1. Load and preprocess MNIST data ( train/test split).
2. Define model architecture:
   • Input layer (784 nodes for 28×28 pixels)
   • Hidden layer (configurable, e.g., 128 neurons)
   • Output layer (10 neurons for digits 0–9)
3. Initialize weights and biases randomly.
4. Implement forward pass:
   • Hidden activations (ReLU)
   • Output layer (softmax)
5. Implement backpropagation:
   • Compute gradients using chain rule
   • Use cross-entropy loss
6. Train with gradient descent, updating weights over epochs, and print accuracy logs (e.g., every 10 iterations)