Thanks both of you for your replies - you really helped me alot! To Peter I must say that this is a very nice piece of code in classify.m. I finally understand, why I had to take all those Linear Algebra lessons in my studies...
Greg Heath wrote:
> use matrix pseudoinversion. I successfully used the > latter for more than 20 years.
This is exactly what I needed, since I dont have much training data. It's mostly close to d (dimension of x). When I use 'linear' discrimination thats normally not a problem since a pooled covariance is used. But with the option 'quadratic' the covariance matrix was often ill conditioned. So I implemented pseudoinversion like you recommended and the results I got were better than with 'linear'!
Well, I guess that nothing beats lots of training data but since its impossible in my case that really helped me out.