NetNews Usenet Archive 1992 #20/NN_1992_20.iso/spool/comp/ai/3338

Internet Message Format | 1992-09-07 | 3.2 KB

Path: sparky!uunet!dtix!darwin.sura.net!wupost!sdd.hp.com!news.cs.indiana.edu!noose.ecn.purdue.edu!sparkyfs.erg.sri.com!paris.erg.sri.com!marie
From: [e-mail] (Marie desJardins)
Newsgroups: comp.ai
Subject: Re: AI Winter
Message-ID: <[e-mail]>
Date: 4 Sep 92 16:56:56 GMT
References: <[e-mail]> <[e-mail]> <[e-mail]> <osborn.715581562@dragon>
Sender: [e-mail]
Reply-To: [e-mail] (Marie desJardins)
Organization: SRI International, Menlo Park, CA
Lines: 49

In article <osborn.715581562@dragon>, [e-mail] (Tom
Osborn) writes:
|>[e-mail] (Marie desJardins) writes:
|>>Since when is machine learning not AI?  I find it hard to believe that's
|>>what you're saying; perhaps you could clarify.  (Possibly your point is that
|>>learning *alone* isn't "complete AI"?  But then neither is knowledge
|>>representation and inference.)
|>
|>He's saying ML != AI. I think that's pretty plain. I take ML as the theory
|>and methods of compressing data into semantically exploitable categorical
|>expressions <woa, adjectival overload!>, ie, to explain data simpler than
|>by listing it. I would (if pressed) take AI to be satisficing within a model
|>- the model usually coming from existent conventional human explanations.

Machine learning researchers generally define machine learning as
"improving performance at some task."  Sometimes this involves compressing/
simplifying/generalizing data.  Some ML researchers focus on data compression
and statistical analysis; others focus on developing cognitive models of human
learning.

|>I don't see ML as a subset of AI (complete AI - not a term I would use).
|>(And I don't see it vice versa either). Even the overlap seems, to me,
|>to be far less than 50% of either field. [The aims may overlap moreso].

I have a bit of trouble with your implied distinction between "the goals of
ML" and "the field of ML".  As for "the goals of ML," learning is *central*
to artificial intelligence.  We cannot build truly intelligent systems unless
they can learn new information and improve their performance.

As for "the field of ML" (the current state of research), I would be inclined
to say that ML is at least as close to addressing and solving its goals as is
the overall field of AI.  Granted, current ML research isn't going to solve its
goals tomorrow; then again, current AI research isn't going to solve
*its* goals tomorrow.

|>The original post (Thrift?) lists various methods of ML which I wouldn't
|>accept as within AI. Even NNs have been given the black spot by many AI
|>Labs and principals. Just because they have become more useful and productive
|>doesn't justify a "prodigal returns" conversion? [I raise a political
spectre].

"various methods of ML" is not the same as "the field of ML".  Additionally,
just as not all AI researchers agree on "what is AI", not all ML researchers
agree on "what is ML."  Some, at least, would not classify all of the methods
listed by the original posting as ML.  Incidentally, the ML and NN communities
are actually rather separate.

Marie desJardins
[e-mail]