Chapter 5
Pooling .
A convolution asks every pixel for its opinion. Pooling holds a vote in each small window and keeps one answer.
Chapter 3 shrank maps with stride: skip positions, and the output gets smaller. Pooling is the other common way to shrink a map, and for a long stretch of CNN history it was the main one. LeNet, AlexNet and VGG all shrink their maps by pooling.
Pooling slides a window over a map, just like a kernel, but it has no weights. It applies a fixed rule to each window instead: keep the largest value, or keep the average. This chapter covers both rules, what each one throws away, and a property of max pooling that often gets overstated: what happens when the input moves by a pixel.
The figures use the same sample image as chapter 3, and every idea from that chapter carries over. Window size plays the role of kernel size, stride means the same thing, and the output size comes from the same formula.
Modules
Each module ends in a short knowledge check. Passing it unlocks the next module. Your progress is saved in this browser.