larry@JPL-VLSI.ARPA (02/08/86)
From: larry@Jpl-VLSI.ARPA
I haven't read the Tech Review article; perhaps I shall just to see how
different will be my interpretation of it from the opinions heard here. The
discussion has made me want to offer some ideas of my own.
What we lump under AI is several different fields of research with often very
different if not contradictory approaches. As a dilletante in the AI field I
perceive the following:
COGNITIVE PSYCHOLOGY (a more restricted area than Cognitive Science) attempts
to understand biologically based thinking using behavioral and psychiatric
concepts and methods. This includes the effects emotional and social forces
exert on cognition. This group is increasingly borrowing from the following
groups.
COGNITIVE SCIENCE attempts to broaden the study to include machine-based
cognition. CS introduces heavy doses of metaphysics, logic, linguistics, and
information theory. My impression is that this area is too heavily invested
in symbol-processing research and could profitably spend more time on analog
computation and associative memories. These may better model humans' near-
instantaneous decision-making, which is more like doing a vector-sum than
doing massively parallel logical inferences.
PATTERN RECOGNITION, ROBOTICS, ETC. attempts to engineer cognition into
machines. Many workers in this field have a strong "hard-science" background
and a pragmatic approach; they often don't care whether they reproduce or
whether they mimic biological cognition.
EXPERT SYSTEMS, KNOWLEDGE ENGINEERING is more software engineering than
hardware engineering. Logic, computer science, and database theory are strong
here. Some of the simpler expert systems are imminently practical and have
been around for decades--though "programmed" into trouble-shooting books and
the like rather than a computer. (And while we're on this, most of what now
passes for rule-based programming could be done in BASIC or assembly language,
including self-modifying code, using fairly simple table-driven techniques.)
And perhaps several more groups could be distinguished. Of course, there are
plenty of exceptions to these categories, but humans do self-select into
groups and distill ideas and techniques into a rudimentary group persona.
If I were to characterize myself, I'd probably say that I'm less interested in
AI than IA--Intelligence Amplification. I'm interested by attempts to create
machine versions of human intelligence and I have little doubt that all the
vaunted "mystical" abilities of humans will eventually be reproduced,
including self-awareness.
Some of these abilities may be much easier to reproduce than we suppose:
intuition, for instance. I'm an artist in several media and use intuition
routinely. I've spent a lot of time introspecting about what happens when I
"solve" artistic problems, and I've learned how to "program" my undermind so
that I can promise solutions with considerable reliability. I believe I could
build an intuitive computer.
But what fascinates me is the idea of building systems which combine the best
capabilities of human and machine to overcome the limits of both. I think
it's much more economical, practical, and probably even humane to, say, make a
language-translation system that uses computers to do rapid, rough transla-
tions of 99% of a text and uses human sensitivities and skills to polish and
validate the translations. (Stated like that it sounds like two batch jobs
with a pipe between them. My concept is an interactive system with both human
and computer collaborating on the job, with the human doing continuous shaping
and scheduling of the entire process.)
Now I'll go back to being an interested by-stander for another six months!
Larry @ JPL-VLSI.arpa