Re: Working on nnet package

[Francesco Faccio]

Julien Bect wrote:

3) Before starting to work on bringing nnet back to life, have you considered other potentially useful ways to spend your time and contribute to OF?  There are many well-developed open source packages for neural networks, and it would be interesting to have bindings for them in Octave, in the form of OF packages.

4) For instance, Søren Hauberg wrote a Octave bindings for FANN [4], which were later modified by Hans Fugal [5].  I don't know how mature this package is.  Some adaptations might be necessary to bring the package up to date with FANN 2.1 and Octave 4.2.0.  Such a package would be an interesting addition to OF.

5) Other popular NN packages worth considering include Torch [6], TensorFlow [7] (Google's NN library), NetLab [8] (a little old perhaps)...

I am not saying that you should give up your project of bringing nnet back to life, if this is what you want to do.  Just that there are other options worth considering...

Hi Julien,

thank you very much for your detailed answer, I would really like to have a thorougly tested Matlab compatible nnet package in Octave.  There are also some colleagues of mine which would be interested to contribute on this project from March/April and it would be nice to get them involved. 

I have already considered to use TensorFlow or Torch as external libraries, which both have a C++ interface, writing the most important functions as oct-files (I've already done something like this in my GSoC project).

Moreover, starting from March, I will work on the development of the Deep Learning part of the package (documented here: [1][2]) and I will make full use of these libraries. 

At the moment I am trying to fix the most simple functions (the ones which do not require to be written as oct-file) and to add tests, then I will start linking libraries.

Do you think it would be interesting for OF to keep Matlab compatibility and to use these libraries?

Thank you,

Francesco 

[1]https://it.mathworks.com/help/nnet/convolutional-neural-networks.html
[2]https://it.mathworks.com/help/nnet/autoencoders.html