Read the video. Skip the scroll.

So let’s start with the simplest possible network — just one neuron. It takes a few numbers in, multiplies each by a weight, adds them up, and squashes the result into something between zero and one.

Now, at the beginning those weights are random. So the answer it gives you is basically garbage, and that’s fine, because the whole point is what happens next.

We compare the output with the answer we wanted and measure how far off we were. That single number is the loss. Everything else is about making it smaller.

Here’s the key trick. Because the whole network is just a chain of simple operations, we can ask: if I nudge this one weight up a tiny bit, does the loss go up or down? And by how much?

Backpropagation is just the bookkeeping that answers that question for every weight at once, starting from the output and working backwards through the layers.

Then we take a small step in the direction that reduces the loss — that’s gradient descent — and we do it again, and again, across thousands of small batches of examples.

After enough of these steps something remarkable happens: the weights settle into values that capture patterns nobody wrote down by hand.