Re: [Dclib-devel] pegasos in dlib

[Alex S. <al...@st...>]

I think I understand the pegasos API pretty well, and I've read the relevant
research papers. One aspect that has been a source of frustration is the
lack of predictability in the quality of the output decision function. As
I'm searching the parameter space for optimal values, I'm finding that very
small changes produce large, random variations in the classification
percentage. Adding or removing features has a similar effect. It seems like
pegasos is too easily trapped in local minima during its optimization, and
it's making poor decisions about what support vectors to start with. The
data I'm working with is extremely noisy, but there is a lot of it (65,000
samples with 50 features). I'd appreciate some insight into how to improve
performance.

Frankly, I'm also a little curious about you as a person. Your developer
page has no information on it. Where do you live? Where do you work?

On Thu, Jul 23, 2009 at 11:42 AM, Davis King <
dav...@us...> wrote:

> Thanks, good to know you find it useful :)   I'm not particular, you can
> email me and the developer list, or post in the help or open discussion
> forms on dlib's sourceforge page.   Whatever floats your boat.
>
> So what's your question?
>
> Cheers,
> Davis