convolutional neural network

A convolutional neural network, or CNN, is the kind of artificial brain that lets computers see. Instead of staring at a whole photo at once, it slides a tiny window — a little filter — across the image, looking at one small patch at a time, the way you might run a magnifying glass over a map. Each filter is tuned to spot one simple thing: a slanting edge, a splash of color, a curve.

The clever part is stacking these layers. The first layer finds plain edges; the next combines edges into corners and textures; the layer above that assembles them into eyes, wheels, or petals; and the top layer finally says "cat" or "stop sign." Built from the ground up, simple parts become rich shapes become whole objects — and the network learns which filters to use by practicing on millions of labeled pictures.

A common worry is that a CNN just memorizes the exact photos it trains on. Mostly it doesn't: because the same small filter is reused across every patch of the image, the network tends to learn patterns that work anywhere — a face is a face whether it sits in the top corner or dead center. Memorizing, or "overfitting," can still happen, which is why training uses huge, varied datasets. But that knack for reusable patterns is exactly why CNNs became the workhorse of computer vision, powering everything from phone cameras to medical scans.