Neurok
A neural network library, from scratch
- 0deep learning frameworks used, NumPy only
- 5→8→4→1network trained on a 2D polynomial surface
Neurok is a compact neural network library built from scratch on NumPy, written to make every step of training explicit rather than hidden behind a framework.
It has fully connected linear layers with He initialization, a mean-squared-error loss, and a training loop with configurable epochs and learning rate. Backpropagation is written out by hand: each layer computes its weight gradient, bias gradient and the gradient it passes upstream.
The included example generates data from a 2D polynomial, engineers polynomial features, trains a 5 → 8 → 4 → 1 network on it, and plots the true and predicted surfaces in 3D.
It deliberately leaves out activation functions, regularization, modern optimizers and GPU support, to keep the core mechanics readable.