How does a machine learn? Explained without a single formula

My mother recently asked me what exactly I study. I said "machine learning" and she made the same face I would make if someone talked to me about commercial law. So I tried another way, and that other way is this article. There are no formulas. There are no strange words. And there is a real experiment planted in the middle of the text so you do not have to take my word for anything.

A machine does not learn like you do. You understand things; it only adjusts numbers. Imagine a box with thousands of little dials. You show it an example ("this email is spam") and the box checks what it would have answered with its current dials. If it gets it wrong, it turns each dial a tiny bit in the direction that would have made the error smaller. Once is nothing. Repeated hundreds of thousands of times, the box ends up answering well. That is, honestly, the whole secret.

The trick is in "the right direction". How does the box know which way to turn each dial? It measures how wrong it was and works out, for every dial, whether turning it right or left shrinks the error. It is like finding the lowest point of a valley blindfolded: you feel the slope with your foot and take a small step downhill. Thousands of small steps later, you are at the bottom. If the steps are too big, you overshoot and bounce from one hillside to the other; too small, and you never arrive.

What you see above is not a video: it is a real, tiny neural network training in your browser right now. The points are the examples and the coloured background is what the network believes about each region. Draw your own points, hit the "Spiral" preset to make it hard, and raise the learning rate — the size of the step — until the network loses its mind. You just broke a training run exactly the way real ones break.

What is this for? Everything we call artificial intelligence today — the translator, your phone's assistant, your bank's fraud detector — is this same idea with bigger boxes: more dials, more examples, more small steps. In my projects I use exactly this mechanism to predict whether a customer will stop paying or to answer questions over documents. The box changes size; the valley trick is always the same.

If this left you wanting more, I have a whole page where the story is told scene by scene as you scroll: How does a machine learn?. And if you dare to try the version with jargon, the rest of the blog is yours.