Chapter 4 · The deep learning era (2006–2016)
Pooling
Pooling shrinks feature maps, keeping the strongest signals and adding tolerance to small shifts.
Chapter 4 · The deep learning era (2006–2016)
Pooling shrinks feature maps, keeping the strongest signals and adding tolerance to small shifts.
Pooling cuts a map into blocks and writes one number for each. Max pooling keeps the biggest, so a strong signal survives even if its exact spot is lost.
A convolution makes a map as large as the picture, and a network makes many of them. Passing all that along is slow, and most of it is not needed.
Drag on the left grid to change the numbers between 0 and 9. Tap a pooled cell to see its block. Then nudge the picture and watch how much changes before and after pooling.
Block 6: 9, 5, 4, 9 → biggest is 9.
A teaching version: whole-number cells, blocks that do not overlap, and edge pixels that shift off the grid are lost. Real networks pool every feature map the same way.
A 2×2 max pool turns the block 1, 7, 3, 2 into…
What does pooling do to the size of a map?
Why does pooling help with small shifts?