Chapter 6
Receptive fields.
Every value in a CNN looks at the image through a window. Stack layers, and the window grows.
In chapter 3, one output read a 3×3 patch of the image. That’s easy to picture when the layer reads the image directly. But the second layer reads the first layer’s output map, not the image, and the third reads the second’s. So which pixels does a value in layer 10 actually depend on?
This chapter answers that question. It covers how to work out the size of that region by hand, why pooling and stride make it grow so fast, and why pixels in its middle count for more than pixels near its edge.
Every figure in this chapter is the same tool: a stack of layers you can build, drawn two ways. One view looks down on the stack from above. The other cuts through it and shows one row of every map from the side. As in chapter 5, the input is the duck, here at 32×32. Every convolution uses same padding, so a convolution at stride 1 keeps its map the same size.
Modules
Each module ends in a short knowledge check. Passing it unlocks the next module. Your progress is saved in this browser.