Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!ncar!noao!asuvax!enuxha!rao
From: rao@enuxha.eas.asu.edu (Arun Rao)
Newsgroups: comp.ai.neural-nets
Subject: Re: request for philosophic reactions to connectionism
Summary: More on ART
Keywords: ART, Grossberg
Message-ID: <124@enuxha.eas.asu.edu>
Date: 23 Apr 89 16:26:40 GMT
References: <370@eurtrx.UUCP| <18496@gatech.edu> <7894@phoenix.Princeton.EDU> <1@bucsb.UUCP>
Organization: Arizona State Univ, Tempe
Lines: 60

In article <1@bucsb.UUCP>, adverb@bucsb.UUCP (Josh Krieger) writes:
> Please stop dismissing ART out of ignorance!
> 
> Grossberg's papers are exceptionally difficult to understand because
> each discovery is dependent on the "minimal" biological building
  [stuff deleted]
> ART is a beautiful discovery.
> 
  [simple explanation deleted]
> 
> The advantages:
> 
>  1) Once a pattern is learned retrieval time occurs in one
>     feedforward pass.
>  2) ART has a vigilance level which allows it to
>     place input patterns into categories based upon course
>     or fine distinctions (Course destinctions would be
>     classifying the letters C and D in the same category, while fine
>     would involve placing them in different categories.
>  3) Old learning is not washed away by the
>     "blooming, buzzing confusion", of the real world.
>     ART is stable to old learning yet plastic to new information.
>  4) ART will converge regardless of the parameter settings.
>     There is no tweaking of learning rates.
>  5) ART has a vast amount of solid psychological and biological
>     evidence for its existence.
> 

	I would not presume to pass judgement on any theory, but I dispute
the (seemingly) widely held view that ART is a novel method of
"learning", as it were, and of classification. I have actually waded through
the kilometers of prose that Grossberg has written, simulated his equations
and otherwise beaten the subject to death. A lot may have still escaped me,
but I have reached the conclusion that ART is nothing but a parallel
clustering algorithm in which exemplars are stored in the (bottom-up)
weights. Effectively, a parallel dot-product is computed between the
unknown vector and all stored exemplars and the maximum dot-product
chosen. It turns out that the dot-product scheme is "imperfect" in the
sense that multiple vectors may generate identical dot-products, and so
the top-down adaptive filter is used to correct the ambiguity thus
produced.

	Conventional clustering methods function in almost exactly the
same manner - if they were parallelized, they would do all that 
ART does, complete with "learning". IMHO, the novelty in ART lies not
so much in the method by which classification and learning is accomplished
(and to say so would be to do a grave injustice to the scores of workers
who have studied clustering methods and parallel algorithms) as in the
way in which it is psycho-physiologically justified.

	Lippmann has pointed out the clustering algorithm analogy in
"An Introduction to Computing with Neural Nets" (IEEE ASSP Magazine,
April 1987), but he does not emphasize it to any great degree.

- Arun
-- 
Arun Rao
ARPANET: rao@enuxha.asu.edu BITNET: agaxr@asuacvax
950 S. Terrace Road, #B324, Tempe, AZ 85281
Phone: (602) 968-1852 (Home) (602) 965-2657 (Office)