rik@blakey.ucsd.edu (Rik Belew) (03/13/91)
I'm sorry to be entering this conversation so late; I've been gone.
It seems that several misconceptions have grown up around our AI-PhD
Genealogy project that I'd like to correct.
USING THE AI-PHD TREE AS PART OF THE IR TASK
First, this AI-PhD family tree was never considered to have any value on
its own. A large part of my work is involved with (AI approaches to)
the problem of free-text information retrieval (IR), the ``library''
problem: how can you describe documents (books, articles, Email
messages, etc.) in a way that let's you find them later, when they are
``relevant'' to some topic. Any of you who have worked in IR know
that its primary methods depend on statistical analyses of {\em word
(token!) frequencies}.
A large part of my own research has been investigating a role for
connectionist learning techniques in a system called AIR. I treat the
browsing behaviors of users as a (reinforcement) feedback signal that
modifies the (connectionist representation of) keywords and documents
so that, over time, AIR is more likely to retrieve documents that
users find relevant. Obviously, what AIR learns depends
critically on {\em who} it is teaching it. For example, we are
getting ready to drop two instantiations of AIR in front of: i)
Computer Scientists and; and ii) Cognitive Scientists here at UCSD.
The ideas is to examine just how the semantics of a term (e.g.,
``neural network'') differ within these two cultures. Note that, for
these purposes, the cultures are being defined in terms of
departmental memberships. (Thus I am not a ``cognitive scientist''
under this definition, even if I think I am!)
This is where the AI-PhD geneaology comes in. It continues to amaze
me just how far word frequency statistics can go to provide semantic
insights, but there are obviously fundamental limits to any method
that relies exclusively on simple word counts. I am interested in the
AI-PhD genealogy exactly as a {\em suppliment} to information derived
from IR analyses and AIR's learning. Theses provide an admittedly
small (but carefully selected) set of keywords. These words will
provide one bridge to the IR side of the system. But the primary
information provided by theses is about {\em authors} and relations
among authors, and a bit about research institutions. The idea is
that when a user asks for documents ``about'' some topic, we can now
reason not only about where those words happen to occur, but also what
groups of people use those words most, where they work, where they
studied, etc. (Note also that this dependence on both connectionist and
symbolic/semantic network technologies creates a host of research
problems, er opportunities.)
Characterizations of disciplines or ``paradigms'' (see below) in terms
of word use, departmental membership, or academic heritage makes
little sense in isolation. But I would argue that a system which {\em
accumulates evidence} can become quite useful.
ACADEMIC HERITAGE AND A PHILOSOPHY OF SCIENCE
Second, I would be the last person to suggest that once this AI-PhD
family tree was built, it will have captured all of what we mean by
``artificial intelligence research.'' For example:
I expect that the Dartmouth ``founding fathers'' provide
(direct) lineage for a {\em minority} of AI PhD's currently working in
the field.
PhD theses, themselves, are probably some of the {\em worst}
examples of research, in AI or elsewhere. By definition, we are
talking about students who are doing some of their first work. They
had {\em better} improve with time!
I am certain that PhD's account for only a fraction of AI
research.
Some of the analyses I am most interested in performing are designed
to test myths such as these. The fact that one of my own advisors has
publically denied paternity highlights my own illegitimacy. (I just
heard a British after dinner speaker announce a new ``morning after''
pill for men: It changes your blood type.) In fact, advantage of
documenting all of the valuable, non-PhD research is to extrapolate
beyond standard publishing activities. That is, I'd like to provide
an argument for the legitmacy of electronic communication (like this
newsgroup) as an important aspect of future Science.
But if we are to make a science of studying Science, we've got to
start with real, hard data. Much previous work in the sociology and
philosophy of Science has wound up as nothing more than a battle among
myths. I am extremely sympathetic to Kuhnian and post-Kuhnian
analyses that highlight the importance of {\em social} aspects of
Science, for example. Further, I worry that many recent AI analyses of
scientific discovery are deficient in this respect. That's another
motivation for this work.
But I need data to prove my theory. ``Paradigms'' are extremely
appealing constructs, but they're also amorphous. For all of its
faults, the AI-PhD tree represents incontrovertiblem syntactic {\em
facts}, just as word frequencies do. Pat Hayes has suggested
(offline) that we ask for ``significant intellectual influences.''
But this is just the sort of introspection I'm trying to avoid. If
you read Hull's \underline{Science as a Process}, or if you're a
parent, you realize that our influences are often not what we intend.
Asking for committee members is a step towards other, non-advisor
influencors, and it is also a matter of record.
The other class of data I am trying to avoid, at this point, is
related to what I think of as the ``star science'' myth: all science
is done by Nobel Prize winners. Closer to home, we've received the
very good idea of checking on all the AAAI Fellows, and we intend to
do that. But such information is misleading if it isn't placed in the
context of the much larger group of people contributing to the field.
{\em Post hoc} evaluations of meritorious conduct are probably very
useful as motivators in Science, but they shouldn't be taken too
literally.
I would argue that Science is primarily about {\em accumulating}
knowledge, even more than about its initial discovery. My primary
argument for interest in this structure is that traditional academic
relationships, as embodied by PhD genealogies, form a sort of
``skeleton'' around which the rest of scientific knowledge coalesces.
Certainly individuals can pursue their own research program, and
corporations can fund extended investigations into areas that are not
represented in academia whatsoever. But it is hard for me to imagine
truly extended research programs that do not involve multiple
``generations'' of investigators; academia is almost certainly the
most successful system for such propagation. When thinking in such
long time scales, AI is a very young field, but the Newell, Simon, et
al. investigation now manifest in the ``SOAR generation'' seems an
excellent example of the value of such longitudinal attention.
--
Richard K. Belew
Computer Science & Engr. Dept. (C-014)
Univ. California -- San Diego
La Jolla, CA 92093