Skip to content
Learn Motion
ExploreHow it worksMembership
Log in
Learn Motion

How AI Sees: A Conceptual Guide to Image and Video Understanding

1What a Machine Actually Sees2From Pixels to Patterns: Features3How Deep Networks Learn to See4Beyond Labels: Locating and Describing What Is Seen5Adding Time: Understanding Video6How These Systems Learn and How We Judge Them
How Deep Networks Learn to See

Pooling: Trading Detail for Tolerance

4 / 5
A feature map is very precise about where a filter fired, and that precision is actually a problem. Shift the object two pixels to the right and every response shifts with it, so a later layer sees a different pattern. Pooling removes that sensitivity. Take each two-by-two block and keep only the largest value. The map gets half as wide and half as tall, and a response at position ten and a response at position eleven now land in the same output cell. Small shifts stop mattering, which is what people mean by spatial invariance. The cost is real, though. You have thrown away exact position, so the network can only say roughly where something was. For naming an object that is fine. For reporting an exact location it is not, and that is why detection and segmentation need different handling of spatial detail.
0:00 / 0:00

A feature map records exactly where a filter fired. That precision is a liability. If the same object shifts two pixels to the right, every response moves two pixels to the right, and a later layer that expects a response at a particular spot sees a different pattern. Pooling fixes this by summarizing a small neighborhood with a single number — most commonly the maximum value in each two-by-two block.

The effect is a coarser map. A two-by-two max pooling step halves the width and height, so a response that was at position ten and a response that was at position eleven both collapse into the same output cell. Small shifts stop mattering. This is what gives the network tolerance to small translations, and it is why the property is often called spatial invariance: the answer stays roughly the same when the input moves a little.

Pooling is not free. Throwing away position information means the network can no longer say precisely where something was, only roughly. That is an acceptable trade for a classifier, which only needs to name the object. It becomes a problem for tasks that must report exact locations, which is why later architectures for detection and segmentation handle spatial detail differently.

Previous4 / 5Next

Learn Motion

Generate a course. Learn it properly.

Operated by Wuhan Daoyin Technology Co., Ltd.

Contact: [email protected]
Privacy PolicyTerms of Service

© 2026 Learn Motion