Before transformers ate everything, Convolutional Neural Networks (CNNs) taught computers to see — and they still power much of image recognition. The trick: build understanding up in layers.
The hierarchy
pixels → [edges] → [textures/shapes] → [parts] → [objects] → "cat"- Convolution slides small filters across the image, each detecting a pattern (a horizontal edge, a curve).
- Early layers find simple features (edges); deeper layers combine them into complex ones (eyes, wheels).
- Pooling shrinks the image between layers, keeping the important bits and adding position-tolerance.
Why convolution is clever
The same filter is reused across the whole image, so a CNN learns "an edge is an edge anywhere" with far fewer weights than a dense net — and it generalizes across position.
Then and now
| CNNs | Vision Transformers (ViT) |
|---|---|
| Built-in spatial priors | Treat image patches as tokens |
| Great, data-efficient | Win at scale; power Multimodal Models |
The 2012 CNN "AlexNet" kicked off the deep-learning boom — see A Brief History of AI. CNNs are also the classic target of Adversarial Examples.
Related: Adversarial Examples · Multimodal Models · How Neural Networks Learn