Date: 16 Feb 1999 22:10:52 GMT
Newsgroups: comp.ai.philosophy,comp.ai
Message-ID: <7acqdc$cn6$1@soapbox.cs.bham.ac.uk>
References: <7a30ev$stc@ux.cs.niu.edu> <36C5198D.42220848@sandpiper.net> <7a4enb$o3@ux.cs.niu.edu> <36C62879.11434042@jhuapl.edu> <7a5j2p$39$3@its.hooked.net> <36C8B120.1F552624@sandpiper.net>
From: Aaron.nospam.Sloman@cs.bham.ac.uk (Aaron Sloman See text for reply address)
Subject: Rickert, Balter, lookup tables and intelligence

In the comp.ai.philosophy newsgroup it seems that Jim Balter and Neil
Rickert are arguing over whether a machine which had a huge lookup table
(HLUT) determining every facet of its behaviour could be distinguished
from one which had human-like internal information processing
capabilities, and could learn and have thoughts, decisions, preferences,
 etc.

Jim Balter <jqb@sandpiper.net> wrote:

> Just what is it that we would never observe in behavior produced
> from a
> HLUT, that would assure us that it wasn't human?

Neil Rickert responded

> The HLUT would be incapable of learning, in the ways that humans
> learn.

[Balter]
> Capacities and ways are not behavior, so that doesn't answer the question.
>
> ....
> ....A HLUT maps states to states.  At one point
> a HLUT is in a state such that it responds to questions about algebra
> as if it knew nothing about algebra.  It then receives inputs,
> side by side with an algebra student.  It transitions into a state such
> that it acts like it knows a little about algebra.  That it doesn't
> learn in the ways that humans learn is irrelevant to your original
> claim about behavior; it *behaves* in the same way.

I suspect that both Balter and Rickert are right (at least partly) but
are talking past each other. (Not uncommon in newsgroups about AI.)

Jim is correct that IF a machine has an appropriate HLUT built in
advance by someone who has worked out what sorts of behaviours are to be
produced in all possible sequences of external environments, then
no amount of *external* observing and testing can distinguish the
machine from one which learns, thinks, takes decisions, etc. for itself.

If they can, then that just proves the lookup table was not designed
correctly.

But Neil is right in saying that there is a difference between a
machine which works out what to do or say and one which merely produces
responses that someone or something else had worked out for it.

However, the difference would not be visible "from outside". It's a
different kind of difference! It's a matter of what's going on in the
virtual machine that produces the behaviour. You cannot infer what's
going on from the behaviour.

(Encrypted output for which you don't have the key is one special case
where this is obviously true. But the problem is far more general.)

If you are trying to understand a complex information processing machine
which you have not designed yourself, the task of finding out how it
works may be wholly impractical. Even if you could examine the internals
of such a machine, checking and measuring the physical mechanisms in
great detail and precision, "decompiling" its circuilts or brain
processes may be possible in principle but extremely difficult in
practice (as brain scientists and software engineers dealing with old
compiled code for which program sources are lost know).

This version of the "other minds" problem is solved for us by evolution:
we are simply programmed to treat other people and animals as having
minds. We don't believe they have minds as a result of *reasoning* from
*evidence* of any kind. We can't help ourselves.

A robot driven by suitably designed HLUT would have the same effect on
us (if such a thing were physically possible.)

The notion of machine controlled entirely by a lookup table presupposes
that all inputs and outputs are discrete. Perhaps this is supported by
quantum physics (e.g. individual photons hit the retina).

It may be that some people object to the idea of a HLUT because they
assume some (or all) interactions with the environment must be analog,
whereas a lookup table maps a discrete states plus a discrete array of
sensory inputs into a discrete array of output signals.

However, even if biological systems are not discrete, it is possible to
make a discrete/digital system approximate, to a very high degree of
accuracy, a purely analog system, e.g. by using digital to analog and
analog to digital converters at its periphery, with a very high sampling
frequency.

But there are deeper objections to the HLUT, some mentioned below.

[balter]
> The hypothetical HLUT is such that all learning that could possibly be
> done within a finite lifetime in all the environments that anyone could
> describe within a finite lifetime with one finite alphabet  is already
> represented as state transitions within the HLUT.

Rickert probably assumes that no finite alphabet could suffice to
characterise human experience and learning. But sufficiently close
approximations may be achievable. Compare the use of fixed precision
floating point numbers to represent real numbers in computers. If you
take enough care you can simulate continuous processes using them, and
this is done every day in scientific and engineering laboratories.

Of course, as Balter acknowledges later, even using a small finite
alphabet, no machine with a large enough HLUT to pass the kinds of tests
envisaged could possibly be built, because a machine with a table that
anticipated *all* the possible external (discretised) contexts that
could occur in a human life time along with *all* the appropriate
responses, would require a larger memory than could fit into the
physical universe.

(Even storing the game tree for chess would require an impossibly big
brain: I think it has been calculated that there are more nodes in
the tree than electrons in the universe. Life is bigger than chess.)

Another, more interesting, reason why the HLUT is impossible is that no
person or machine could work out *in advance* everything that needs to
be in such a table to accommodate a human lifetime. E.g. it would
involve anticipating all possible cultural and technological
developments in a lifetime, all possible future scientific theories that
might be developed in a life time (including false theories that become
fashionable!), all possible artistic creations to which the individual
might be exposed, and so on.

Being able to anticipate all that would require being able to run some
sort of simulation of a large portion of the universe much faster than
real time. However, if the future after time T included someone running
such a simulation then the person running the simulation just before
time T would have to go even faster. An infinite regress of faster and
faster simulations might be required?

But quite apart from that difficulty it seems to be impossible to gain
all the fine-grained knowledge required to anticipate everything years
ahead. (Maybe that's why Turing restricted his test to 5 minutes!)

Just as a designer could not anticipate all relevant future states,
neither could evolution. I.e., although it may be able to use "trial and
error" learning to work out a set of behaviours adequate to a range of
conditions likely to be met by bacteria or ants, through a period of
exploratory evolution, it could not do the same for the environments
likely to be encountered by humans, and perhaps not chimps and other
various other more or less intelligent species.

One reason for this is that after we develop new capabilities we then
change our environments in such a way as to produce quite novel
situations that were not available in advance for the development of
appropriate behaviours in those situations! (E.g. building grand pianos,
sky-scrapers and computers.)

Instead evolution somehow found a better trick: it developed mechanisms
capable of doing planning and problem solving at "run time", i.e.
allowing the individuals to work things out for themselves, by
constructing new types of behaviours. This requires a specialised
architecture supporting "what if" reasoning capabilities.

This is denied by some people who argue for systems built *entirely* out
of large 1collections of pre-stored reactive behaviours. (That may work
for an ant, as far as I know. But ants don't seem to develop new
scientific theories, design and build new kinds of machines, develop new
languages for controlling machines, prove and use new theorems, etc.
etc.)

Anyhow,

[Balter]
> One might say that the HLUT
> doesn't learn at all, it only *behaves* as if it does.  But the difference,
> to the degree that one can talk about a difference between a class of
> real entities and a class of strictly theoretical, unrealizable, entities,
> is strictly metaphysical; the HLUT extended in space and the
                                     ^^^^^^^^^^^^^^^^^
> real learner extended in time are isomorphic.
               ^^^^^^^^^^^^^^^^

I am not sure that the word "isomorphic" is useful here.

Actually, it is not even correct: The real learner has potential to
follow many branches of action which never occur. But those branches are
not explicitly realised in its brain or in its action: instead he/she/it
has generative mechanisms for *working out* that they should be
performed if needed. (I presume that's part of Rickert's point.)

By contrast the HLUT must have all those possibilities explicitly
represented in advance, e.g. all the sentences that might possibly have
to be uttered in all situations that could occur.
(Unless the designer of the HLUT could work out a *unique* time-line for
everything in the environment, which I suspect is impossible for many
different reasons.)

So it would be more accurate to say that the HLUT corresponds to the
real learner extended in *all possible futures* from the day of its
birth. But it's not clear what follows from that correspondence in this
sort of debate. Calling it an isomorphism doesn't deny the important
differences.

[Balter]
> Real learners undergo state
> changes as they receive inputs; only one state is physically represented at a
> time.

That's just one view of the architecture. It would be an accurate
characterisation of the HLUT, and of some monolithic AI programs
and neural nets.

However, I think it is more illuminating to think of human minds and
brains as collection of concurrent interacting subsystems whose states
change concurrently and not necessarily in any synchronised way. This is
more illuminating because it leads to better explanatory theories, and
is more likely to lead to the design of artificial systems that are like
humans.

(I accept that this comment probably doesn't contradict what Balter
*intended* to say. My own work, developing the idea of a complex
"nearly modular" virtual machine architecture with concurrent
interacting components can be found in papers in the Birmingham
Cognition and Affect FTP directory

    ftp://ftp.cs.bham.ac.uk/pub/groups/cog_affect/0-INDEX.html

Papers near the top of that list try to show, for example, how three
different kinds of emotions, in humans, i.e. primary, secondary and
tertiary emotions, depend on different concurrently active
sub-architectures which evolved at different times, and are shared with
different subsets of animals.)

[Balter]
> HLUTs switch from state to state as they receive inputs; all states
> are physically represented at once.  The only way to be able to build a HLUT
> is to produce a 3-dimensional projection of our 4-dimensional world.
> Since we can't do that in practice, they are only relevant to fantasyland.

Not jsut a projection of our 4-D world, or any particular 4-D world, but
an indefinitely large branching set of *possible* 4-D worlds.

Understanding all the differences between a table incorporating such a
projection and a mechanism that is capable of *generating* that sort of
projection is central to understanding the nature of intelligence.

Even if we agree with Rickert that HLUTs won't do because they don't
work things out for themselves, or learn, it's not at all obvious what
sorts of architectures *are* capable of producing the kinds of "working
out" and learning that go on in human minds (our virtual machines) or
brains (our implementation engines), nor what goes on in other animals.

Brain science, psychology and AI are still a long way from clear,
accurate and deep explanatory theories: there are only fragments.

We don't know how to put them together, nor whether putting them
together would suffice. We could be missing something very important
that nobody can imagine just now. (Penrose thinks it's quantum-gravity
effects, but I think that if there are such things they operate at the
wrong architectural level to answer our questions: discovering them
would be analogous to discovering transistors.)

But that doesn't mean we won't continually deepen our understanding if
we work at it (though some people find that thought distasteful in the
same sort of way that many people found distasteful the thought that we
are not at the centre of the universe, or that we are products of a
biological evolutionary process).

Aaron
===