Introduction
After a convolution, a feature map tells you where a pattern appears, pixel by pixel. Often you don't need that much detail: knowing that "there's a vertical edge somewhere in this 2 × 2 area" is enough. Pooling layers shrink feature maps by summarizing small regions with a single number.
Think of describing a city map to someone. You don't list every house; you say "this neighbourhood has a park, that one has shops". Pooling does the same: each small neighbourhood of the feature map is summarized, so later layers work with a smaller, more general picture.
Pooling has no learnable weights. It simply applies a fixed operation (take the maximum, or the average) to each window.