1
1

There's an old theorem by Cover on the number of linear dichotomies of a set of m points in "general position" in d-dimensional space. Does anyone know if there's been any variants of that which, say, require bounded norm of coefficients of the hyperplanes or require a minimum margin on the hyperplanes? Or, failing there being theorems, are there practical algorithms for estimating the number of bounded norm or minimum margin separated dichotomies for a particular data set one has at hand?

asked Aug 26 '10 at 13:25

Dave%20Lewis's gravatar image

Dave Lewis
755142342

edited Aug 26 '10 at 20:19

Just curious, do you have a reference for that theorem?

(Aug 27 '10 at 04:28) Yaroslav Bulatov

Here's the original reference:

Cover, T. M. (1965). "Geometrical and Statistical Properties of Systems of Linear Inequalities with Applications in Pattern Recognition," IEEE Trans. Elec. Comp., EC-14, 326-334.

It's discussed in various textbooks on machine learning, including

Massoud, M. H. Fundamentals of Artificial Neural Networks (MIT Press, 1995)

The relevant section is:

http://neuron.eng.wayne.edu/tarek/MITbook/chap1/1_3.html

(Aug 27 '10 at 08:29) Dave Lewis

Are you talking about VC dimension?

(Aug 27 '10 at 10:07) Alexandre Passos ♦

Yes and no. What I'm wondering (to jump ahead) is whether having a particular data set at hand can give you tighter bounds than VC dimension would.

(Aug 27 '10 at 10:19) Dave Lewis

@Dave Lewis : In that case, you should be looking at data dependent bounds such as Radamacher complexity.

(Aug 27 '10 at 12:22) spinxl39

@spinxl39 - Thanks, indeed Radamacher complexity is the kind of thing I was looking for. If you want to turn your comment into an answer, I'll accept it as the answer.

(Dec 06 '10 at 11:57) Dave Lewis
showing 5 of 6 show all
Be the first one to answer this question!
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.