In week 1 we have the introduction to Pooling layers. I’m curious if anyone knows why there’s no padding right before the pooling operation.
I’m asking because it seems to me that the “corner pixel problem” that motivates pooling after convolution is present during pooling (i.e. the pixels throughout the middle of the image/matrix are part of “many passes” of the filter while the corner ones just once).
Andrew mentions padding is rarely used during pooling but, for the reasoning above, I’d think it would be used there too.
I understand the reasoning: “…you want to shrink your dimensions in a smart way, which is achieved by Pooling… You could imagine that by taking the max over nearby activations, you still retain the information about the presence of a particular feature in this region, while losing information about its exact location”.
This makes sense but still it doesn’t explain to me it isn’t a problem that the high activations that are max pooled in the middle of the input matrix will “indicate the presence” of the same feature over the “slidings” of the filter that overlap where this high activation exists, whereas a high activation value in a corner will be “picked up” by the max operation only once.