Built for images

Before CNNs, getting a computer to "see" was a losing battle of hand-designed rules. Convolutional Neural Networks (CNNs) changed that by building in one powerful idea: scan the image with small filters that detect local patterns. A filter is a tiny window of weights that slides across the whole image, lighting up wherever it finds its pattern — an edge, a color blob, a texture. Crucially, the same filter is reused everywhere, so a feature is detected no matter where it appears.

Hierarchy, again

CNNs stack these filter layers, and the same feature hierarchy from earlier emerges naturally:

• Early layers detect edges and simple gradients.
• Middle layers combine edges into textures and parts — a corner, an eye, a wheel.
• Deep layers assemble parts into objects — a face, a cat, a car.

Between layers, pooling shrinks the image, keeping the important signal while making the network efficient and more tolerant of small shifts.

Why this design is so efficient

Two properties make CNNs brilliant for vision. Translation invariance — a cat is a cat in any corner of the frame, because filters scan everywhere. And parameter sharing — reusing one small filter across the whole image needs far fewer weights than connecting every pixel to every neuron, so CNNs train efficiently even on large images.

Where they're used

CNNs power image classification, object detection, facial recognition, medical imaging, and self-driving perception. Even as transformers push into vision, CNNs remain a fast, strong, and widely deployed choice — and the clearest way to build intuition for how deep networks turn raw pixels into meaning.

A CNN sees by sliding pattern-detectors across an image and stacking them into a hierarchy — edges to parts to objects. The same trick that finds an edge anywhere finds a face anywhere.

Try this: Look at a photo and consciously decompose it the way a CNN would — first the strong edges, then the textures and parts, then the whole objects. That layered way of seeing is exactly what the network learns, and it makes the architecture intuitive.