Neural Network from Scratch with NumPy

A deep dive into the fundamentals of deep learning, building a neural network from the ground up using only NumPy.

Project Overview

This project focuses on implementing a neural network from scratch using only NumPy. Every functionality—from forward propagation to backpropagation, activation functions, and gradient descent—was developed using NumPy arrays. The goal was to understand and implement the core mathematics and logic behind neural networks without relying on high-level libraries.

Key Results

• 3-layer Model: Achieved 86% accuracy on the MNIST dataset.
• 4-layer Model: Achieved 91% accuracy on the MNIST dataset.

Features

• Completely Custom Implementation: Every part of the neural network, including matrix operations, loss functions, and optimizers, is built using NumPy.
• Focus on Fundamentals: Gained in-depth knowledge of how neural networks work at the most basic level.
• Efficient Training: Achieved high accuracy through careful design and optimization.
• Learning-Centric Approach: Each line of code contributes to a deeper understanding of neural network mechanics.

Technology Used

• Python: Programming language for implementation.
• NumPy: For efficient array operations and mathematical computations.
• MNIST Dataset: Used for training and testing the models.