I was looking with great interest at the new Deeplearning4J node and I was wondering how difficult would be within the knime environment to build a deep neurnet autoencoders which to start with learns molecular fingerprints (e.g. ECFP4) and output molecular fingerprints. Similarly for example to what they have done in this paper:
I am not an expert in encoding/decoding layers so any advice would be very much appreciated. I did not find any autoencoders example workflows on the Knime website.
I had a quick look at the paper you mentioned. They seem to be using Generative Adversarial Autoencoders which are not included in the DL4J Integration, but we have normal Autoencoders in KNIME. However, there was a bug in the DL4J Library for quite some time which prevented the Autoencoders from learning anything. For the new release we updated the library so this could be solved but I couldn't verify that myself yet.
sorry for the late answer. Adverserial networks are quite different to a normal Autoencoder or Deep Belief Network. I'm not sure if they can be used to achieve similar results than in the paper (I did not have time to delve deep into the matter). Also I don't know if DL4J want to add that. I would suggest you could try to find an implementation of a Generative Adversarial Autoencoders in Python. Then you could use the Python Scripting Nodes if you want to use it in KNIME.