[mlpack] Implementing support vector machines in mlpack

[Ryan Curtin]

On Mon, Feb 02, 2015 at 12:46:21AM +0530, Dhawal Arora wrote:
> Hi,
> 
> I am a third year undergraduate doing b.tech computer science in India.I
> have done a few introductory courses in machine learning. Been going
> through the mailing list and have compiled mlpack successfully and worked
> on a few algorithms as per the tutorials. Now i would like to contribute to
> mlpack and also want to get involved with mlpack for gsoc 2015.
> 
> Was going through the issue #396.Have read about manifold learning
> algorithms and want to start working on this issue. If someone could give
> any feedback on this.

Hi Dhawal,

#396 looks pretty open-ended to me (maybe Stephen can comment if he had
more specific ideas than Laplacian Eigenmaps, LLE, and IsoMap).  He's
right that it would be a lot easier if Armadillo had an eigs_pair()
function, kind of like the eig_pair() function [1] but for recovering
only a couple of eigenvectors.  I'll get that implemented hopefully
sometime next week.

If you have a manifold learning technique you're interested in, you
could start implementing that and we can go from there.  As long as the
implementation is high-quality, well-documented, well-tested, and fast,
we can work it into the mlpack code.

> Also regarding implementations, mlpack is not supporting SVM as of now. Is
> it appropriate to work on this for gsoc 2015? Is it too big as a project or
> we can start working on some basic implementations of linear and gaussian
> kernel, later we will cover some more areas. I have a few more ideas on
> this.

You are correct that mlpack currently doesn't have an SVM implementation
and it could certainly use one.  We have kernels in
src/mlpack/core/kernels/, so if you were to write an implementation, it
should use templates to support any type of kernel.  If you are
interested in pursuing SVMs, be sure that (like I wrote above) the
implementation is high-quality, well-documented, well-tested, and fast.
We can help with any of those four things, of course. :)

> Just wanted to know that is mlpack having any tentative list of ideas as of
> now for gsoc 2015?

Yeah, we need to make one soon, but it hasn't happened yet.  Check again
in a week or two...

Also keep in mind that we haven't been accepted into GSoC at this point
(organization applications aren't even open yet), so there is always a
possibility that mlpack will not be selected to participate in GSoC
2015.

Hope this is helpful... let me know if I can clarify anything.

Thanks!

Ryan

-- 
Ryan Curtin    | "You know, I think he's got a point."
ryan at ratml.org |   - The Mayor