|
I am trying to build a convolutional auto-encoder following the method described in this paper for pre-training a deep convolutional neural network. The problem that I am facing is on how to build the decoding (reconstruction) layer. The encoding layer has a convolution step followed by a max-pooling step. In the decoding layer, I need to have some step which approximately reverses the effect of max-pooling i.e. some sort of up-sampling before I perform full convolution to complete the reconstruction (decoding). How method would be best to use for this step? In case you can suggest any other way to build the convolutional auto encoder, kindly suggest. |
|
Not sure if you saw this recent paper that showed up on the archive, : "Visualizing and Understanding Convolutional Neural Networks" It gives a very nice explanation of how to do this. |
This is previously answered here http://metaoptimize.com/qa/questions/12873/how-to-visualize-high-level-features-in-convolutional-deep-networks-with-pooling
I was trying to implement some thing similar. In context of this paper, is it like, when using max-pooling you do the exact same thing as before other than forcing the non-maximal hidden-layer activations to be 0. This forces the hidden-layer representation to be sparse, but the decoder part does not have to change.
Yes, that is how it would turnout to be.