When reading some deep learning papers, which sometimes mentioned that max-pooling layer for downsampling can also be used for increasing the number of feature channels(maps). This confused me a lot. It looks to me the max-pooling layer can down sample the size, but should keep the number of original feature maps.
$\begingroup$
$\endgroup$
2
-
3$\begingroup$ Can you provide a link to a paper where this was stated for context? $\endgroup$– Tarin ZiyaeeSep 22, 2016 at 6:36
-
$\begingroup$ @TarinZiyaee I came here because I try to understand AlexNet. $\endgroup$– Felix CrazzolaraMar 1, 2021 at 11:19
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
2
As you say, Max-Pooling does not modify the number of channels, it only downsamples in the spatial dimensions.
But, as the spatial dimensions are smaller, given a constant computational budget, you could increase the number of channels in subsequent layers without increasing the computational load (up to a limit). This could be what those papers meant.
-
$\begingroup$ I don't really get it. I see what you are saying about a constant computational budget however I do not see how, if the programmer were in inclined to do so, they would increase the number of features in the max pooling layer. I thought the max pooling layer only looks at the feature maps of the previous conv. layer. If there were eg 60 feature maps in the conv. layer, how could simply doing max pooling on this conv. layer leave you with any more feature maps than the conv. layer you are doing pooling on initially had? $\endgroup$– laraOct 31, 2016 at 4:11
-
$\begingroup$ @LaraJordan I didn't say that Max-Pooling does this, only that you can design a network that takes advantage of the downsampling done by Max-Pooling. It is just a design decision. $\endgroup$ Oct 31, 2016 at 9:27
$\begingroup$
$\endgroup$
https://arxiv.org/pdf/1902.11107v2.pdf
channel max pooling is max pooling along the channels You do operation similar to max pooling on the channel side, click on link
