[mlpack-git] [mlpack] Investigate implementing SGD for LRSDP (#395)
notifications at github.com
Tue Feb 10 14:45:17 EST 2015
This is sort of an open ended proposal. The point is, I'm not really satisfied w/ any of the SDP methods I worked on for mlpack-- we have LR-SDP which requires lots of hand-holding in practice, and we have this new primal-dual interior point method which, while theoretically sound, is quite slow in practice. If we could have something that (a) restricts the class of problems it solves but (b) scales well in practice and (c) has theoretical convergence guarantees, then that'd be awesome.
However, I don't want to just keep adding SDP solvers to mlpack (they impose maintenance burden) unless there is a clear win. Hence, it'd be cool to have an implementation of alecton, but I'd be hesitant to merge it unless it actually worked well. Since this is a newer algorithm, there is more risk of this happening. I want to warn you ahead of time, so that if you do decide to work on it and it doesn't work well, there won't be any hard feelings :smile:
If you're deterred by this, but still want to work on something related to SDPs in mlpack, I have a few "lower risk" ideas related to improving our existing solvers. Let me know.
Reply to this email directly or view it on GitHub:
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the mlpack-git