|
Recently SVMs have experienced an explosion of optimization algorithms for multiple kernel learning and linear SVMs. The latter in particular is around the point where it can be scaled to almost any size of data. Unfortunately, I hven't been able to find any good survey that can take me through the underpinnings of modern optimization techniques for SVMs. Can anyone recommend a survey? |
|
As far as I know one still needs to be written. The research doesn't seem to be over, as there are still serious issues with nonlinear SVMs, parallel optimization, and when should one optimize on the primal and when should one optimize on the dual. |
|
I think for linear SVMs Kristian Woodsend' PhD thesis and for MKL, the JMLR paper "Lp-norm Multiple Kernel Learning" have simply covered and classified last researches. |
|
You can look at http://mitpress.mit.edu/books/chapters/0262026252chap1.pdf |
It turns out LibLinear has a lot of great papers and code on its website. Not everything I'd hoped for, but still good. http://www.csie.ntu.edu.tw/~cjlin/liblinear/