Pooling
Squashing a whole sequence of vectors into a single vector that stands for the entire input.
Think of it like
Like summarizing a long meeting into one sentence — you collapse many points into one takeaway.
Example
To classify a review, a model pools the per-word vectors into one “whole-review” vector before the classification head.
How it actually works
Networks produce one vector per token, but many tasks need a single representation of the entire input. Pooling combines them — by averaging, taking the max, or reading a special summary token. The method matters: averaging is robust but blurry, while a dedicated summary token can learn to focus on what counts. It’s the bridge from per-token detail to one whole-input meaning.
For product teams
The step that turns “many word-vectors” into one thing you can score or search on.
For engineers
An aggregation over token representations (mean/max/CLS) producing a fixed-size vector.
Related
- Mean Pooling — The most common averaging variant.
- CLS Token — A dedicated summary-token approach.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome