[mlpack-git] [mlpack] improve speed of SparseAutoencoder and make it more flexible (#451)

[stereomatchingkiss]

>(1) we can improve the speed of the SparseAutoencoder

Yeah, I would like to do that, and intent to provide gpu version in the future.

>(2) we can reuse the ANN code and remove virtually all of the existing code for SparseAutoencoder

I would like to resue the ANN code + existing code for SparseAutoencoder if possible:)

>do you think we could use the ANN code entirely to represent the sparse autoencoder

I think it would not be able to do that, I wrote the reasons [at here!](https://github.com/mlpack/mlpack/pull/451#issuecomment-150390803). Either the layer of ANN need to do some changes and apply type_traits + static_assert to tell the users "this layer is a bad choice" or create a new folder to collect meaningful layer for autoencoder(but this would generate duplicate codes).

>do you think we could use the ANN code entirely to represent the sparse autoencoder

About the FNN part, I am not sure it could be resue or not, but it looks a lot similar with current implementation, I will need to dig into the codes, run some examples to find out the results are correct or not and compare the speed with old implementation. 

If it works, then we maybe able to provide autoencoder and sparse autoencoeder at the same time.

> I'm not a particular expert on autoencoders

Me neither, I am not even an expert of machine learning, I only know some basic ideas of it and need some decent tools to help me finish image recognition/classification tasks. At first I intent to write a library for myself, but later on I find out this is impractical and prefer to join an existing open source project:).

In short, I think expert like zoq would know the direction better.

---
Reply to this email directly or view it on GitHub:
https://github.com/mlpack/mlpack/pull/451#issuecomment-161498521
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack-git/attachments/20151202/e0226905/attachment.html>