CNNs: built to exploit the structure of images
Convolutional neural networks, or CNNs, are designed around a key insight about images: a pattern like an edge or a texture can appear anywhere in the frame, so instead of learning a completely separate detector for every possible position, a CNN slides small filters across the image, reusing the same learned weights at every location. This is called parameter sharing, and it dramatically reduces the number of weights needed compared to connecting every pixel to every neuron, while also making the network naturally good at recognizing a pattern regardless of where it appears in the image.
This design made CNNs the dominant architecture for computer vision for years, powering things like medical image analysis, self-driving car perception systems, and photo tagging. They're still widely used today, especially in resource-constrained settings like mobile devices, though vision transformers have increasingly challenged them in state-of-the-art image benchmarks.