Week 1 Assignment 1 Backpropagation | Max Pooling mask for multiple maximum values in A_prev

[Woreviam]

Hi everyone, I’m Marco from Peru.
I understood the idea to create those mask to backward the propagation of dA to dA_prev (before the maxpool). The assignment made us to assume that there will be only one maximum value in each patch of A_prev for that exercise, but what about the case in which we have multiple maximum values in some patches?, how can we distribute our dA elements to each correspoding dA_prev patch with multiple maximun value in its A_prev patch?

I was thinking maybe all the corresponding maximum values positions for a patch in A_prev got its values in dA_prev modified in those positions by the same factor. Seems to be understandable, and easy to implement, but the fact that the assignment made us to not consider these cases made me doubt.

Do guys know additional sources about this process?

Thanks in advance.

Marco

[paulinpaloalto]

What they mean by ‘you don’t need to consider’ is that it is handled automatically by the way you create the mask: you compare the matrix against the maximum value and in the case that there are multiple equal maximum values, the mask will have more than one position that is equal to 1, right?

So they’re not saying we don’t handle that case: they’re just saying that it’s easy.