I was reading the ZF Net paper and it used the term Deconvnet on some searching it seems this is the wrong term and rather we use transposed convolutions instead. I understood how transposed convolutions work but I still don't get how does this show the information stored in a particular layer or the features captured by a particular layer
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
I went through the ZFNet paper and it seems that Deconvnet is different than transposed convolution. However, the idea of both is quite similar and it is easy to get confused.
First, let's be clear about transposed convolution. There are many recent blogs and papers about it and here is one nicely written blog: https://towardsdatascience.com/what-is-transposed-convolutional-layer-40e5e6e31c11).
But, the ZFNet paper is old and it cites this paper https://ieeexplore.ieee.org/document/6126474 for Deconvnet idea. The idea is similar to transposed convolution in a sense that a feature map with smaller dimensions is transformed into a feature map of bigger dimensions. It has three steps Unpooling, Rectification, and Filtering which makes it different than that of transposed convolution. To understand the Deconvnet idea I recommend doing a much more in-depth analysis of the ZFNet paper. You may find these slides helpful http://cs.nyu.edu/~fergus/drafts/utexas2.pdf
So, why the confusion? Deconvolution is originally defined as stated on this Wikipedia page: https://en.wikipedia.org/wiki/Deconvolution. And as convolution does the exact opposite, so probably the upsampling of the feature map (transforming lower-dimensional feature maps into higher-dimensional feature maps) could be considered as a deconvolution operation. However, after reading several resources I conclude that there are several ways in which one can upsample the feature maps, like pixel shuffle (https://nico-curti.github.io/NumPyNet/NumPyNet/layers/pixelshuffle_layer.html) and transposed convolution (discussed earlier). Thus, all these operations do the same high-level thing but in a very different way (resulting in even different results).
I also recommend going through this thread for further reading: What are deconvolutional layers?
