[mlpack] Algorithm Optimization proposal Query

[Ryan Curtin]

On Sat, Mar 16, 2019 at 07:37:05PM +0200, Rana Khamis wrote:
> Hello,
> 
> My name is Rana Mohammed Khamis , a computer science junior student in Ain
> Shams University who is interested in Machine learning.
> 
> For Algorithm Optimization Project I have chosen Linear Regression, my
> question is would it be a good choice to use parallelization using openmp
> to enhance the speed needed for the Algorithm , since i'm gonna make a
> report after running the benchmark and compare the results of both, i
> expect that parallelization if done accurately it would make a 25%
> enhancement.

Hi Rana,

Thanks for getting in touch.  I'm glad to see you are interested in the
algorithm optimization project, but there are a couple issues:

 - Linear regression is a really simple algorithm and I would not expect
   it to take very long at all (maybe a week maximum) to parallelize it.
   So I think that unless you are proposing some radical new strategies
   for linear regression, this wouldn't be enough work for a GSoC
   project.  Maybe you might consider looking into other algorithms?

 - For linear regression in particular, the part that takes a long time
   is the call to arma::solve().  That in turn will be a call to a BLAS
   or LAPACK library, which is most likely to be OpenBLAS---which is
   already parallelized.  So I think that there is not much that can be
   gained for the optimization of linear regression specifically.

However, it may still be an interesting exploration to see what kind of
speedup you can get with linear regression, and looking into that might
help you understand the codebase a bit better and prepare a better
proposal.

I hope these comments are helpful; let me know if I can clarify
anything.

Thanks!

Ryan

-- 
Ryan Curtin    | "This is how Number One works!"
ryan at ratml.org |   - Number One