=7 cellspacing=0 width=100% bgcolor=ccccff> « back to results for ""
Below is a cache of http://www.cogs.indiana.edu/andy/AIandFacesofReason.pdf. It's a snapshot of the page taken as our search engine crawled the Web.
The web site itself may have changed. You can check the current page or check for previous versions at the Internet Archive. Yahoo! is not affiliated with the authors of this page or responsible for its content.
Artificial Intelligence and Artificial Intelligence and the Many Faces of Reason² (Refereed book
chapter) in S. Stich and T. Warfield (eds) THE BLACKWELL GUIDE TO
PHILOSOPHY OF MIND (BLACKWELL, 2003)







Artificial Intelligence and
The Many Faces of Reason


Andy Clark
Philosophy/Neuroscience/Psychology Program
Department of Philosophy
Campus Box 1073
Washington University in St. Louis
St. Louis, MO 63130
USA
andy@twinearth.wustl.edu
















Note: Correspondence address after June 1
st
, 2000 will be

School of Cognitive and Computing Sciences
University of Sussex
Brighton
BN1
9QH
England
UNITED KINGDOM

The old email address will remain active.
2
0.
Pulling A Thread

I shall focus this discussion on one small thread in the increasingly complex weave of
Artificial Intelligence and Philosophy of Mind: the attempt to explain how rational
thought is mechanically possible. This is, historically, the crucial place where Artificial
Intelligence meets Philosophy of Mind. But it is, I shall argue, a place in flux. For our
conceptions of what rational thought and reason are, and of what kinds of mechanism
might explain them, are in a state of transition. To get a sense of this sea change, I shall
compare several visions and approaches, starting with what might be termed the Turing-
Fodor conception of mechanical reason, proceeding through connectionism with its skill-
based model of reason, then moving to issues arising from robotics, neuroscientific
studies of emotion and reason, and work on ecological rationality. As we shall see
there is probably both more, and less, to human rationality than originally met the eye.

First, though, the basic (and I do mean basic) story

1.
The Core Idea, Classically Morphed

One core idea, common to all the approaches Ill consider today, is that sometimes form
can do duty for meaning. This is surely the central insight upon which all attempts to
give a mechanical account of reason are based. Broadly understood, it is this same trick
that is at work in logic, in the Turing Machine, in symbolic Artificial Intelligence, in
connectionist artificial intelligence, and even in anti-representationalist robotics. The
trick is to organize and orchestrate some set of non-semantically specifiable properties or
features so that a device thus built, in a suitable environment, can end up displaying
semantic good behavior. The term semantic good behaviors covers, intentionally, a
wide variety of things. It covers the capacity to carry out deductive inferences, to make
3
good guesses, to behave appropriately upon receipt of an input or stimulus, and so on.
Anything that (crudely put) looks like it knows what it is doing, is exhibiting semantic
good behavior: cases include the logician who infers A from (C
?
- A,C), the person who
chooses to take out an umbrella because they believe it will rain and desire to stay dry,
the dog who chooses the food rather that the toxin, the robot that recovers its balance and
keeps on walking after one leg is damaged. Theres a lot of semantic good behavior
around, and we understand some of it a whole lot better than the rest. Where, though,
does reason come into the picture?

Reason-governed behavior is, arguably at least, a special subset of what I am calling
semantic good behavior. It is Jerry Fodors view, for example, that is was not until the
work of Turing that we began to have a sense of how rationality (which Ill assume to
mean reason-governed behavior) could be mechanically possible (for a nice capsule
statement, see Fodor (1998 p. 204 205)). Formal logic showed us that truth
preservation could be ensured simply by attending to form, not meaning. B follows from
A & B regardless of what A means and what B means, and if your keep to rules defined
over the shapes of symbols and connectives you will never infer a falsehood from true
premises, even if you have no idea what either the premises or the conclusions are about.
Turing, as Fodor notes, showed that for all such formally (by shape) specifiable
routines, a well-programmed machine could replace the human.
It is at about this point that what was initially just an assertion of physicalist faith (that
somehow or other, semantic good behavior has always and everywhere an explanatorily
sufficient material base) morphs into a genuine research program targeting reason-
governed behavior. The idea, rapidly enshrined in the research program of classical,
symbolic Artificial Intelligence, was that reason could be mechanically explained as the
operation of appropriate computational processes on symbols, where symbols are non-
4
semantically individable items (items typed by form, shape, voltage, whatever) and
computational processes are mechanical, automatic processes that recognize, write and
amend symbols in accordance with rules (which themselves, up to a certain point, can be
expressed as symbols). In such systems, as Haugeland (1981, p. 23) famously remarks,
if you take care of the syntax [the non-semantic features and properties] the semantics
will take care of itself. The core idea, as viewed through the lens of both Turings
remarkable achievements and then further developments in classical Artificial
Intelligence, thus began to look both more concrete, and less general. It became the idea,
in Fodors words, that:
some, at least, of what makes minds rational is their ability to perform
computations on thoughts; when thoughtsare assumed to be syntactically
structured, and where computation means formal operations in the manner of
Turing
Fodor (1998) p.205.
The general idea of using form (broadly construed) to do duty for meaning, thus gently
morphed into the Turing Machine dominated vision of reading, writing and transposing
symbols: a vision which found full expression in early work in Artificial Intelligence.
Here we encounter Newell and Simons (1976) depiction of intelligence as grounded in
the operations of so-called physical symbol systems: systems in which non-semantically
identifiable entities act as the vehicles of specific contents (thus becoming symbols)
and are subject to a variety of familiar operations (typically copying, combining, creating
and destroying the symbols, according to instructions). For example, the story
understanding program of Schank (1975) used a special event description language to
encode the kind of background knowledge needed to respond sensibly to questions about
5
simple stories, thus developing a symbolic data-base to help it fill in the missing
details.
Considered as stories about how rational, reason-guided thought is mechanically
possible, the classical approach thus displays a satisfying directness. It explains
semantically sensible thought-transitions (they enjoyed the meal, so they probably left a
tip, its raining, I hate the rain, so Ill take an umbrella) by imagining that each
participating thought has an inner symbolic echo, and that these inner echoes share
relevant aspects of the structure of the thought. As a result, syntax-sensitive processes
can regulate processes of inference (thought-to-thought transitions) in ways that respect
semantic relations between the thoughts.

2.
The Core Idea, Non-classically Morphed
The idea that reason-guided thought transitions are grounded in syntactically driven
operations on inner symbol strings has a famous competitor. The competing idea,
favored by (many) researchers working with artificial neural networks, is that reason-
guided thought-transitions are grounded in the vector-to-vector transformations supported
by a parallel web of simple processing elements. A proper expression of the full details
of this contrast is beyond the scope of this paper (see Clark (1989)(1993) for my best
attempts). But we can at least note one especially relevant point of (I think) genuine
contrast. It concerns what Ill call the best targets of the two approaches. For classical
(Turing Machine-like) Artificial Intelligence, the best targets are rational inferences that
can be displayed and modeled in sentential space. By sentential space I mean an
abstract space populated by meaning-carrying structures (interpreted syntactic items) that
share the logical form of sentences: sequential strings of meaningful elements, in which
different kinds of syntactic item reliably stand for different things, and in which the
6
overall meaning is a function of the items (tokens) and their sequential order, including
the modifying effects of other tokens (e.g. the not in it is no