Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!seismo!think!craig
From: craig@think.COM (Craig Stanfill)
Newsgroups: net.ai,net.cog-eng
Subject: Re: transition from AI to Cognitive Science (was: Re: Notes on AAAI '86)
Message-ID: <6158@think.COM>
Date: Sat, 6-Sep-86 15:09:31 EDT
Article-I.D.: think.6158
Posted: Sat Sep  6 15:09:31 1986
Date-Received: Sat, 6-Sep-86 20:37:11 EDT
References: <959@hounx.UUCP> <963@batcomputer.TN.CORNELL.EDU> <7602@tekecs.UUCP>
Reply-To: craig@godot.think.com.UUCP (Craig Stanfill)
Organization: Thinking Machines, Cambridge, MA
Lines: 19
Xref: mnetor net.ai:1113 net.cog-eng:268

> I find it very interesting that there is so much excitement generated over
> parallel processing computer systems by the AI community.  Interesting in
> that the problems of AI (the intractability of: language, vision, and general
> cognition to name a few) are not anywhere near limited by computational
> power but by our lack of understanding. [...]

For the last year, I have been working on AI on the Connection Machine,
which is a massively parallel computer.  Depending on the application,
the CM is between 100 and 1000 times faster than a Symbolics 36xx.  I
have performed some experiments on models of reasoning from memory
(Memory Based Reasoning, Stannfill and Waltz, TMC Technical Report).
Some of these experiments required 5 hours on a 32,000 processor CM.  I,
for one, do not consider a 500-5000 hour experiment on a Symbolics a
practical way to work.

More substantially, having a massively parallel machine changes the way
you think about writing programs.  When certain operations become 1000
times faster, what you put into the inner loop of a program may change
drasticly.