This paper analyzes max pooling and average pooling, as it is used in many convolutional neural networks (CNNs).
## Why pooling is used
* invariance to image transformations
* more compact representations (- remove irrelevant information)
* better robustness to noise and clutter
## Max pooling or average pooling?
No clear answer to that. Sometimes one seems to be better, sometimes the other, sometimes something in between.