🚀 go-pugleaf

wombat@cs.montana.edu (Izurieta) writes:


>	I have been reading a few papers in segmentation of environments
>into regions.  Apparently, this is done by humans to create cognitive
>maps of some areas.  These regions in turn, form part of a hierarchy.

>	Here is my question,  What are the advantages of using a hierarchy
>to subdivide an environment?  Are there any computational advantages?

I am most interested about this topic, but from other perspective.
In 1989 I developed such an algorithm to train Self-Organizing
Feature Maps in logaritmic time (O(N) is reduced to O(log N), where
N is the number of units (neurons)). At that time the only advantage
was speed, but I believe that there should be other (representitional)
advantages as well. So I would appreciate if you can point out
the references  you were talking about (I have already Hinton's
article in Scientific American).

The only (currently available) reference to my work was in Neural Net
conference 1990 (San Diego), available also via ftp from:

lut.fi (157.24.10.8)
account: ftp
password: --see instructions
directory: reports/neural-nets
file: selfhier.ps.Z

Since then we have used this net a lot in our research, and 
future papers will show very similar results that mentioned
in Hintons article. (Our work is mainly in computer vision and
artificial intelligence, using neural networks).

>				Clem
>				(wombat@cs.montana.edu)

Pasi Koikkalainen                    (permanent address)
 ]bo Akademi ( - 22.Sept.93)         Lappeenranta Univ. of Tech.
 Inst. Informationsbehandling        P.O.BOX 20, 53851-LPR
 20520 Turku, Finland                Finland
 email: pkoikkal@ra.abo.fi           email: pako@lut.fi