[mlpack] Need guidance for parallel implementation of ICA

[Ryan Curtin]

On Tue, Feb 10, 2015 at 01:14:04PM +0530, shashank gupta wrote:
> Hello,
> 
> I am Shashank Gupta, Masters student at BITS Pilani. I am working on
> parallel implementation of Independent Component Analysis as part of
> College project involving parallel implementation of data mining
> algorithms.
> 
> I came across Radical algorithm implemented in mlpack. It's results were
> promising and better relative to fastICA or infomax , mostly for which
> parallel implementation is available.
> 
> I am thinking to implement Radical algorithm given in mlpack on MPI. But I
> am not sure how to proceed.
> 
> Can you tell me how should I proceed with it.

Hi Shashank,

I recently worked with a student who tried to parallelize RADICAL, but
the project did not result in any working code.  She was planning on
using OpenMP; so that might be a place to start.  Her strategy was to
parallelize the Jacobian rotations, since there are (I think) d^2 of
those.  RADICAL performs particularly slowly in high dimensions, so I
think that could be accelerated via parallelization.

Please let me (or the mailing list) know how your project goes.  If you
find serial improvements to the code, I'm more than happy to incorporate
them.  At this time, though, an MPI implementation wouldn't mesh nicely
with the existing codebase (in the long term, that should change).

Thanks,

Ryan

-- 
Ryan Curtin    | "Long live the new flesh."
ryan at ratml.org |   - Max