Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!mailrus!uwm.edu!ux1.cso.uiuc.edu!ux1.cso.uiuc.edu!m.cs.uiuc.edu!s.cs.uiuc.edu!mehra
From: mehra@s.cs.uiuc.edu
Newsgroups: comp.ai.neural-nets
Subject: Re: Minimum # of internal nodes to form
Message-ID: <218600003@s.cs.uiuc.edu>
Date: 17 Oct 89 16:40:09 GMT
References: <50402@<1989Oct12>
Lines: 18
Nf-ID: #R:<1989Oct12:50402:s.cs.uiuc.edu:218600003:000:714
Nf-From: s.cs.uiuc.edu!mehra    Oct 16 20:19:00 1989


># Abu-Mostafa and Eric Baum also have interesting results in this direction
># but I won't get into that here.
>
>Can you please post the references

Abu-Mostafa's paper on "Random problems" suggests using the difference between
Kolmogorov Complexity and Shannon Entropy as a measure of "hidden complexity".

Baum and Haussler's paper is called "What size net gives valid generalization?"
They bring in several ideas from computational learning theory, including
the relationship between the VCdimension of a 1-hidden-layer n/w of LTUs, and
the number of examples it can "shatter".

You will probably find both papers in NIPS proceedings. I am not prepared to
answer any further questions on this matter.

Pankaj