Chapter 8
Upsampling and transposed convolution.
Every chapter so far has made maps smaller. This one makes them bigger again, and finds out what that costs.
Stride and pooling shrink a map so that each value can see more of the image. For a network that gives one answer per picture, that’s the whole story: shrink until a few numbers are left, then decide. But some networks have to answer for every pixel, like chapter 7’s network that marked which pixels were duck. Those networks shrink the map and then have to grow it back.
This chapter covers the two ways to grow a map. The first is a fixed rule with no weights: copy each value, or blend its neighbors. The second is a layer with weights, the transposed convolution, and it’s the backward step of chapter 7 put to work as a layer of its own. Writing a convolution as a matrix shows where the name comes from. The chapter ends with a flaw the transposed convolution is known for, checkerboard artifacts, and the usual fix.
Modules
Each module ends in a short knowledge check. Passing it unlocks the next module. Your progress is saved in this browser.