Convolutional Networks

Why images need a different kind of layer than plain fully-connected networks, and what a convolution intuitively does.

A small filter slides across the image, looking for the same pattern everywhere.

A fully-connected layer applied to a 224×224 image would need a separate weight for every pixel-to-neuron connection — over 100,000 weights per output neuron, with no notion that nearby pixels are related. Convolutional layers fix this by sliding a small learned filter (say 3×3 pixels) across the entire image, reusing the same weights everywhere.

The key insight: local patterns repeat

An edge, a curve, an eye — these visual patterns can appear anywhere in an image. A convolution learns one small filter that detects 'an edge' and applies it at every location, rather than learning a separate 'edge detector' for the top-left corner and a different one for the bottom-right.

Raw pixels

Conv layer 1

detects edges, colors, simple textures

Conv layer 2

combines edges into shapes, corners

Conv layer 3+

combines shapes into object parts

Classifier head

combines parts into 'cat', 'dog', etc.

CNNs build a visual hierarchy: simple features compose into increasingly abstract ones.

This is exactly the architecture behind image classifiers, object detectors, and — before transformers took over — most of computer vision.

The key insight: local patterns repeat

Check your understanding