You're talking about provenance (colour), but
that's irrelevant to
my point, which is that the man in the
chinese room (or rather, the system of
man + rules) is
indistinguishable from another system (i.e. a native
speaker)
whom you concede *does* process semantics. You
state that some of the rules
are non-algorithmic, presumably
to argue that a computer in the room can't
duplicate those
rules, ergo can never process semantics, but in fact
"discretion" is rather easy to mimic using weighted
probabilities (which can
take into account consistency with
past choices, and various other heuristics).
So I can write
an algorithmic program to give to a computer inside the
chinese room, and you won't know whether you're talking to a
computer or to a
human native speaker of Chinese.
The real reason the man-in-the-room
scenario is difficult is
not that the man (or computer) is incapable of
learning
semantics, it's that he's deprived of learning
opportunities. The
rulebook might tell him that an
acceptable response to [the chinese characters
meaning] "I
love you" is "I love you too" and that a frequent response
to "I
hate you" is "I hate you too", but it would be very
difficult to know the
difference between love and hate. In
contrast, a Chinese child would be
afforded the opportunity
to associate each of those phrases with very different
facial expressions, tone of voice, etc. Give a computer
(with suitable
hardware e.g. cameras, microphones) those
same clues, and it could "do
semantics" just as well as any
child.
The man-in-the-room scenario may
sound a little like Plato's
Cave or its more recent
mind-trapped-inside-the-skull
variations (including the Matrix movies) but
it's
different. Those are about whether a mind can trust its
inputs, and
therefore whether a philosopher can ever be
confident of what's "real". The
man-in-the-room scenario
is more like the dilemma of Egyptologists before the
Rosetta
stone. The man in the room may never learn Chinese
semantics,
because the scenario is designed to prevent him
from ever encountering
semantically useful information. It
says nothing at all about whether
computers can learn
semantics, it just says that computers can't learn
semantics
if you don't give them any semantic information.
Whether the
man inside the room has any semantic information
is not the same question as
whether the system of man+rules
has any semantic information. The rules would
encode some
associations from which you could deduce a few things -
certain
words appearing together frequently could mean they
are semantically related,
for example. If you knew some
grammatical principles from other languages, you
might be
able to guess which words were verbs and which were
conjunctions.
Also, to be convincing, the rules would have
to contain some sort of
personality model, which might be
mined for a few more crumbs. But in a
stringent application
of the Chinese Room scenario, the rules could be
carefully
constructed to exclude any information about whether a
certain
Chinese character means "love" or "hate", "but" or
"and", "yes" or "no".
Thus the system of man + rules clearly does not "know"
Chinese semantics,
at least not more than fractionally.
I'm changing my mind: I now see why
it's fair to say that
the Chinese Room system (man PLUS rules) is faking it,
not
really "doing semantics", even if you can't tell the
difference by looking
at the system's output. But the
Chinese Room example has nothing to do with
the capabilities
of computers - it's simply a very artificial situation.
Limit a computer's input, and there's a limit to what it can
learn.
Your
previous post presumes that discretion was in the
rules; if it's in the rules
it's in the system, thus the
system can "do discretion" which is not a terribly
interesting result to me.
You seem to think the system can only work with a
man
(not a computer) inside the room, because discretion is
something
computers can't do. Maybe I'll think about this
some more and conclude that
computers can only do fake
discretion, or rather maybe I'll conclude that the
difference between fake discretion and real discretion is
meaningful in some
way. Putting that aside for now, I think
you'll concede that fake discretion
is good enough to get
the system to work convincingly.
This is getting
far afield for your article, and I'm not
sure how much of this is responsive to
your last post, but
it's interesting.
So to start heading back to where we
started: the Chinese
Room system (man plus rules) implements ersatz semantics,
whoe outputs happen to be indistinguishable from real
semantics. You could,
indisputably, write a computer program
that did ersatz semantics; for example a
program could
translate between human languages without having any
semantic
references except from one language to the other.
(I've written part of
something similar myself; come to
think of it, many computer-science students
have written
something similar: a compiler. It never occured to me that
this
was something less than real semantics.) I still hold
that an actual computer
system (including programming and
training) is capable of real semantics.
You may object that computers are never given semantic
information, just
bits which it must process
algorithmically. I respond that the human brain is
never
given semantic information either, just neurotransmitters
and action
potentials. (Plato's Cave...) It's possible that
it's processing these signals
in some non-algorithmic way,
but I don't see any evidence for that. Semantics
aren't all
that special, just an association between words and
"meaning".
"Meaning" is a funny word, since meanings can be
imaginary ("unicorn",
"leprechaun"), proving that many
meanings are simply associations with other
words in the
language, but ultimately the idea of "semantics" is that
some
meanings are references to information not contained in
syntax. Supply that
information to a well-programmed
computer, and (real, not ersatz) semantics are
a piece of
cake.
Totally unrelated question: suppose a computer program
does
not explicitly give a procedure for accomplishing a task;
instead, the
program is an algorithm for learning, and to
apply it to a specific task, you
need a specific set of
training data. (Or, the program is a process for
applying a
learned model, and to apply it to a specific task, you need
a data
file containing a specific model.)
Under US law, can you patent the program+data
combination
for each specific task?
[ Reply to This | Parent | # ]
|