Where backprop shows up

Here's the quietly astonishing thing: almost every trained neural network on Earth learns by the exact algorithm you just built. Different data, different architectures, wildly different tasks — but underneath, the same forward-pass / backward-pass / take-a-step loop. This closing tour shows the range.

One algorithm under all of it

Every other course on this site describes a network that had to be trained — and this is how. The CNN that sees, the transformer that reads, the diffusion model that paints, the embeddings that capture meaning: all of them started as random weights and were shaped by backpropagation and gradient descent. It's the closest thing modern AI has to a universal law.

It's also strikingly old: the algorithm was popularized in the 1980s. What changed wasn't the math but the scale — huge datasets and the GPUs to run those two passes billions of times.

That's the course

Learning, demystified end to end:

• A loss turns "wrong" into one number to minimize.
• The gradient is the slope telling each weight which way is downhill.
• The chain rule finds a deep weight's slope by multiplying local ones.
• The backward pass computes all those gradients in one efficient sweep.
• And gradient descent takes the step, over and over.

So the next time you read that a model was "trained," you'll know the machinery exactly: guess, measure the error, send it backward, nudge every weight — a billion times over.

If you enjoyed this, the other courses show what all this training produces — from how ChatGPT works to how images are generated.