difficulty 3/5
Pooling
Replacing each window of a feature map with a single summary — usually its maximum — to shrink the grid and make the representation less sensitive to exactly where a feature sat.
Intuition
A convolution tells you where it found an edge, to the pixel. Usually you do not want that. Whether the photograph contains a cat should not depend on the cat being two pixels left of where it was in training, and a representation that changes when it moves is carrying precision nobody asked for: "objects hardly ever occur exactly at the same place" [1, §7.5].
Pooling throws that precision away deliberately. Take a small window, keep one number — the largest, usually — and move on. The question "was there a strong edge somewhere around here?" survives; "exactly which pixel?" does not.
The side effect is arithmetic and just as important: a 2 × 2 window at stride 2 produces a quarter as many activations, so every layer above it is four times cheaper.
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/
- Understanding the Effective Receptive Field in Deep Convolutional Neural Networks (2017)paperhttps://arxiv.org/abs/1701.04128