radford@ai.toronto.edu (Radford Neal) (06/03/90)
The Chinese Room discussion shows dangerous signs of reappearing. In the
belief that just ignoring it may not work, I'm posting this in the hopes
of at least turning it in a direction that hasn't already been beaten to
death.
One version of what the Chinese Room is all about can be phrased as the
question, "Is the passing the Turing Test an adequate criterion for
intelligence?"
Usually, one might describe the Turing Test as:
If the program answers numerous probing questions in a manner
that can't be distinguished from a human's answers, then we
should consider that it thinks.
This is wrong, however. What we really mean by the Turing Test is:
If we are convinced that the program will in the future answer
probing questions in a manner that can't be distinguished
from a human's answers, then we should also be convinced that
it thinks.
This makes explicit that we are making an inductive conclusion here - a
jump from finite evidence to a general conclusion. This can make a
big difference, given the sort of unrealistic thought experiments that
arise in this discussion.
For starters, consider the occasional objection that the program might
be performing table lookup - just spitting out canned answers to
the questions with no real "thought" involved. The usual objection
to this is that the required table of answers to all concievable questions
would require an amount of memory greater than the number of atoms
in the known universe, but let's ignore this as being a mere quibble.
The objection is still invalid. We have to ask, "how was the table
filled in?" If we _know_ for certain that the table entries were produced
by a random number generator, then we will never make the inductive
leap to the conclusion that future questions will be answered sensibly.
No matter how many good answers were produced in the past, if the
table was randomly generated the next answer is likely to be garbage.
The program doesn't pass the Turing Test, so the objection has no force.
On the other hand, we might have reason to think that the table is
non-random. Maybe the programmer (a very long-lived fellow) filled
in the table with sensible answers to all possible questions, e.g.
with the answers _he_ would have produced. We don't have to be certain
of this to start, just a suspicion is enough. When the program gives
lots of good answers, this suspicion becomes a near certainty and we
become convinced that future answers will also be good. The program
passes the Turing Test.
But wait a minute... What we really mean is that whatever entity we
are interviewing when we ask the questions passes the Turing Test. It
is clear, however, that the real causal explanation for why the answers
are good is that the programmer put sensible answers in the table. We
are really interviewing the programmer via a bizzare time-delay process.
The Turing Test has just demonstrated that the programmer thinks.
I think similar objections apply to "more realistic" scenarios. Remember
that to pass the Turing Test, the program has to give reasonable
answers to questions such as "What does a plucked chicken look like
before it's been roasted?". Now really, as software engineers, how might
we tackle this problem? Surely not by including conventional symbolic
rules for constructing English replies to English questions of this sort!
Instead, we'll simply include a few million digitized images of everyday
scenes, and have the program construct replies based on this data, together
with general English comprehension/generation rules.
Including digitized images doesn't violate the description of such
systems as Searle's Chinese Room, but somehow I think it's not quite
what Searle imagines. It is now clear that there is a causal
connection between actual chickens and the replies to the questions.
Indeed, once again, if we had reason to rule out such a causal
connection, the program would not pass the Turing Test, because we
would decline to make the inductive conclusion that it would answer
future questions about chickens reasonably.
I think such considerations break down Harnad's whole distinction between
the ordinary Turing Test and the "Total Turing Test", and refute the kurfuffle
about "symbol grounding". Any program that passes the Turing Test will
necessarily be constructed via a process that establishes a causal link
between features of the actual world and the sentences output by the
program that refer to such features.
Radford Neal