I work with someone who has developed a deep learning neural net topic model. It appears to work better than standard approaches such as Latent Dirichlet Allocation. Given Mathematica's set of functions that facilitate building a neural net I wonder if it would make sense to convert it in Mathematica, or would it just be better to wrap the python code and implement it in that form?
My experience in building neural nets in Mathematica is limited enough that I don't think I'm up for the project, but if anyone is interested in doing the conversion, I can send you to the code. I think a topic model capability would make a powerful addition to the Mathematica library.
Of course generally there is interest, but you would need to provide more details about your net. Take a look at what is already built in Wolfram, you might already see it there: