Backpropagation in a pool layer (Subsamplig layer) in CNN

Question

My doubt is how I can return the error at the pool level, because when I calculate the derivative, there is only 1 element in 4 (for example, when using the 2x2 union kernel) that affects the result of the forward.

Answer 1

Suppose you have a four-element matrix M

ab cd 

and maxpool (M) returns d . Then the maxpool function really depends only on d . So, the derivative of maxpool with respect to d is 1, and its derivative with respect to a, b, c is zero. So you backpropagate 1 to the module corresponding to d , and you skip back zero for other units.