Comment un réseau de neurones apprend

Picture a cook trying to make a soup you’ll love. The first bowl tastes wrong. What should change? A pinch more salt, less pepper? The cook needs to notice the difference between the soup they made and the soup they wanted, then try again. That’s the challenge.

A neural network learns in a similar way. Give it a picture and ask, “Cat or dog?” It passes the picture through layers of connected units. Each connection has a setting that affects the signals passing through. Together, those settings produce a guess. At first, that guess can be wildly wrong.

Now show it the right answer. The network checks how wrong its guess was. Then it sends that error backward through the layers to find how each setting affected the guess. It nudges the settings, like the cook tweaking the recipe after tasting. One nudge is small; many examples gradually improve its guesses.

Think of a phone sorting photos of cats and dogs. It wasn’t born knowing what makes them different; it learned from labeled pictures and corrections. Each new example tests its current settings, and each mistake can help tune them. So how does a neural network learn? It adjusts connections after comparing guesses with correct answers.