Pooling
also: max pooling · average pooling
Downsampling a feature map (usually taking the max or average over small regions) to shrink it and add invariance.
Pooling downsamples a feature map by summarizing small regions — max pooling takes the largest value in each window, average pooling the mean. It shrinks the spatial size (less compute downstream) and adds a bit of translation invariance, since the exact position within the window stops mattering.
Worked example: 2×2 max pooling turns a 4×4 feature map into 2×2 by keeping the strongest activation in each quadrant — halving each dimension and keeping the most salient signal. Gotcha: pooling discards precise location, which helps classification (a cat is a cat wherever it is) but hurts tasks needing exact position (segmentation, detection) — which is why some modern architectures replace pooling with strided convolutions, and vision transformers drop it entirely in favor of attention.