A CNN (Convolutional Neural Network) processes grid-like data such as images by sliding small learnable filters across the input to detect local patterns — edges, then textures, then shapes — building hierarchy layer by layer. Weight sharing (the same filter everywhere) makes it efficient and translation-invariant.
Worked example: an early layer’s filters fire on edges anywhere in the image; deeper layers combine those into eyes, wheels, letters — so a CNN recognizes a cat wherever it appears without a separate detector per position. Gotcha: CNNs bake in a strong prior (locality and translation invariance) that makes them data-efficient for images, but that same prior limits them where global relationships matter — which is why vision transformers, with less built-in prior but more flexibility, overtake CNNs given enough data.