Re: what is SVM? what are the good features of SVM?

> From: Paul Victor Birke ([EMAIL PROTECTED])
>
> As an aside, what is the current status of SVM.
> One of the criticisms I heard a while back was the
> need for a lot of support vectors to get reasonable
> answers.

Paul,

your message didn't make it to my news server, and I found it later on
google - sorry for a late response.

I'm not an expert in SVM, but judging from what comes out at machine
learning conferences, these are the basic directions:
* the use of optimization that's more efficient than quadratic
programming on large datasets (LS-SVM (least squares SVM), reduced
SVM, proximal SVM, ...)
* a probabilistic perspective is needed to properly deal with
uncertainty (Bayes point machines, relevance vector machines,
probability estimates from SVMs)
* kernel issues (kernel parameter tuning (very important, BTW), kernel
learning, kernels for all sorts of problems)
* stability and smoothing issues (regularization)
* transduction (use of unlabelled data for supervised learning)
* applications
* all sorts of bounds and bounds on bounds and bounds on
approximations of bounds and approximate bounds on approximations of
bounds, fortunately nobody's been spotted doing lower bounds of upper
bounds, and unfortunately nobody's ever been spotted studying lower
bounds on the utility of bounds.

The references I give are scanty and incomplete, but I hope it helped.

Aleks