difficulty 3/5
Convolutional neural network
A stack that alternates convolution and downsampling so that features grow from edges to objects while the grid shrinks — the architecture that made image models tractable.
Intuition
A convolutional network is not one idea, it is one idea repeated: look at a small neighbourhood, summarise, shrink, look again. What changes as you go up is what "a small neighbourhood" contains. At the bottom it is pixels, so the layer can only find edges. Two blocks up it is a neighbourhood of edge detectors, so the layer can find corners and textures. Higher still, it is a neighbourhood of those: in D2L's account the lowest layers "might come to detect edges, colors, and textures", and higher ones parts and then whole objects [1, §8.1.1].
Nothing in the architecture says "learn edges first". The hierarchy falls out of the fact that each layer can only see the layer below, and that layer has already been reduced.
The other half of the design is economics. The convolutional prior — the same weights everywhere, looking only locally — is what makes a model of an image affordable at all, and the worked example below puts a number on how much.
Formal definition
How it works
Where it's used
Why & when
Worked example
Pitfalls
Check yourself
3 questions
The questions on this page come with access. They are graded on the server, so the answers stay there until you have given yours.
Sources
- Dive into Deep Learningbook · retrieved 2026-09-05https://d2l.ai/
- Deep Residual Learning for Image Recognition (2015)paperhttps://arxiv.org/abs/1512.03385
- Deep Learningbookhttps://www.deeplearningbook.org/contents/convnets.html