Hi David.
Thanks for your email. I would appreciate it if we could keep the
discussion on the list for as long as possible,
as other people might have something to say or are interested.

I recently posted a gist: https://gist.github.com/2061456
And there is also a branch by me:
https://github.com/amueller/scikit-learn/tree/multilayer_perceptron

These would be good places to start.

This implements SGD.
My biggest wish would be to do the SGD in Cython, which
should result in a major speedup. This can probably
be done using the recently used helper classes from
the linear SGD, which might need to be extended.

As far as deep learning goes: I agree, this is to much
and I think it is at the moment not in the scope of the
scikit.

Apart from writing the SGD, there should be support for
differnt loss-functions. In my current implementation, the
hinge-loss support is not working yet.

I feel it hard to judge the amount of work necessary
since I am not so experienced in Cython myself, but
I think it should be possible to extent SGD to other optimization
methods. Levenberg-Marquardt is definitely interesting
and would be good to have.
Maybe you can compare it with resilient backpropagation
and also include that if it proves helpful.

I would like to hear from the "SGD people" what they
think about this project and what they think is doable in the time.

Cheers,
Andy

ps: I would be interested in hearing what parts of sklearn your
advisor uses :)


On 03/19/2012 01:57 PM, David Marek wrote:
> Hi Andreas,
>
> My name is David Marek. I am studying for Master's Degree in
> Theoretical Computer Science at Faculty of Mathematics and Physics,
> Charles University in Prague. I would like to apply for gsoc and work
> on scikit-learn. I like the Neural Networks idea. I have experience in
> using different learning algorithms in Matlab and I have written
> simple backpropagation algorithm.
>
> What do you think the soc project should contain? My idea is to create
> one estimator for feed forward neural networks classification and one
> for regression with different methods available for learning. One is
> the stochastic gradient descent algorithm. However, I am not sure that
> project would be big enough. Maybe add another learning algorithms,
> e.g. some second order learning algorithm (Levenberg-Marquardt is
> quite popular in Matlab although it isn't very usable on large
> dataset). What do you think? On the other hand there would be a lot of
> work on documentation and creating new examples and tests. I am not
> sure how much time would this project take.
>
> If I should specify what my idea doesn't contain, it's deep learning.
> I think that's beyond the scope of this idea.
>
> I would really like to work on scikit-learn because I am probably
> going to use it in my Master's thesis which will be about statistical
> methods for spoken dialogue management. I found scikit-learn when my
> advisor told me he uses it for his research in machine learning.
> However, I am not sure I am going to be able to participate in gsoc
> right now, because I have also applied for google internship and
> passed all interviews but haven't received any information about
> available positions yet. I sent my application too late so I am not
> sure there will be any place available for me and in that case I would
> like to do gsoc. I have been told I should get the final verdict
> before 26th March which means I should know if I'll be able to
> participate before the student application opens. So that's a little
> warning.
>
> Thanks
>
> David Marek
>    


------------------------------------------------------------------------------
This SF email is sponsosred by:
Try Windows Azure free for 90 days Click Here 
http://p.sf.net/sfu/sfd2d-msazure
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to