Neural Network from Scratch
A professional implementation of a neural network using only NumPy for MNIST digit classification.
- Status
- Finished and delivered. It does what it set out to do and no more work is planned.
- Kind
- Machine Learning
- Started
- Jun 2025

About it
The problem
Frameworks like TensorFlow and PyTorch make neural networks easy to use but hide how they work, so many learners never see the mathematics behind backpropagation and optimisation.
The approach
I implemented a complete neural network in NumPy alone, with a modular object-oriented design that makes every component swappable and inspectable.
- Six activation functions, five loss functions and four weight initialisation methods.
- SGD with momentum, learning-rate scheduling and early stopping, controlled by more than 20 CLI parameters.
- A drawing interface for live predictions, plus reports with accuracy curves, loss graphs and confusion matrices.
The result
The network reaches 98.06% test accuracy on MNIST, competitive with framework-based baselines, and serves as both a learning tool and a base for custom experiments.
What it does3
Pure NumPy Implementation
High Accuracy
Comprehensive CLI
More work

Rone Arena API
A public FastAPI for Mobile Legends hero data, analytics and player tools, with a web playground and SDKs.

BeliMadu.com
An online honey shop on Next.js 16 and Cloudflare Workers, with a WhatsApp-checkout cart and a Vitest suite.

ridwaanhall.com
This site: a portfolio and blog on Next.js 16 with an admin, Markdown versions of every page and a CV generated from the About page.
Comments0
Signed-in readers can comment.