a hands-on course

A Neural Net
From Scratch

(Backpropagation in NumPy)

Build and train a neural network with nothing but NumPy — and finally see what PyTorch's autograd hides.

A free, hands-on tutorial: build a neural network from scratch in pure NumPy — no framework, no autograd. Implement the forward pass, the loss and backpropagation by hand, then train it to solve a problem a single layer can't. The under-the-hood companion to the PyTorch course. Runs entirely on your laptop.

in the real world

Writing backprop once, by hand, is the fastest way to truly understand how every neural network learns — and to stop treating PyTorch and TensorFlow as magic.

Start with the setup →

Part 1 — The network & forward pass

1

Setup & a problem to solve

Install NumPy, then generate a toy dataset that a straight line provably cannot separate.

2

The forward pass

Wire up weights, a hidden layer with ReLU, and a sigmoid output — a network that predicts, badly, from random weights.

Part 2 — Loss & backpropagation

3

Measuring how wrong

Binary cross-entropy: one number that says how far the network's probabilities are from the truth.

4

Backpropagation by hand

The chain rule, in code: compute the gradient of the loss for every weight, layer by layer, backwards.

5

The training loop

Forward, loss, backward, nudge the weights — repeat. Watch the loss fall and accuracy climb to ~100%.

Part 3 — Train it and see it learn

6

See it, and what's next

Plot the curved boundary the network learned — then see that your backward() is exactly what autograd automates.