[This file may be converted to HTML later]

From Aaron Sloman Wed Jan 22 05:04:51 GMT 2003
To: PSYCHE-B@LISTSERV.UH.EDU
Subject: Re: Memories are made of this (and a virtual machine question)

Stan aimed a question at me. It's answer is quite complex. I don't
know if I can make it clear and convincing, or even just clear,
but I'll try.

> Speaking of metaphysical issues I have one for Aaron. What evidence
> do you have that virtual machines tell us about the qualia (feels)
> aspect of consciousness. I think that is the aspect that we have been
> discussing.

Individual virtual machines produced by this or that software
engineer won't tell us anything specific about qualia or
consciousness.

Likewise a virtual machine that proves algebraic theorems will
probably not tell you much about a virtual machine that controls a
chemical plant or an aeroplane's automatic landing system or one
that paints pictures, like the AARON program of Harold Cohen. (You
can download it or watch it paint from a distance here
http://www.kurzweilcyberart.com/ )

However, learning about virtual machines, including how to design
them, debug them, document them, extend them, criticise them,
exposes you to powerful new ways of thinking about processes,
mechanisms and causation. We can then combine these new ideas and
techniques with information gained from lots of other disciplines.

Of course, ideas concerning virtual machines are not completely
new. Freud's id, ego and superego were of course virtual machines
not physical machines. (Perhaps he thought some aspects of the id
were physical: I don't know.)

Long before Freud there were philosophers, logicians and others
who speculated about relations between physical and non-physical
machines.

For instance Aristotle thought of souls (minds) as somehow
intimately linked to the body: the soul is the form of the body,
but not all bodies had souls of the same kind, since only some
could do reasoning, have beliefs, etc.
    http://azaz.essortment.com/aristotlesoulp_rmeb.htm

In between we get Shakespeare saying that there is no art to find
the mind's construction the face, which is true in general, even
though he allowed Lady Macbeth to chide her husband for letting
others see his mind (especially his guilt) in his face
    'Your face my thane is as a book, where men may read strange
     matters'

And when the great psychologist wrote:
    Love is not love
    which alters when it alteration finds

he was not talking about brains (what did Shakespeare know about the
amygdala?) but about a virtual machine in which emotional states and
attitudes can sometimes be changed on the basis of new information
(finding alteration in a person can change your attitude to that person)
even though there are some virtual machine states that are not so easily
dislodged, e.g. love -- like the poor sods who go on loving the one who
beats them up.

(Powerful attractors in a state space? But it will have to be a state
space for a machine whose states include semantic contents. )

There have also been economists, historians, anthropologists, etc. who
study social, economic and political virtual machines and their
interactions. I guess entomologists are another case.

So, as someone else pointed out, people have been studying virtual
machines of many kinds for centuries past.

What is new (in the last half century) is that whereas

- previously people were studying 'from the outside' very complex
  systems whose character was very hard to understand and whose
  relationship to the underlying physical reality was a source of
  many unsolved puzzles (e.g. about how non-physical causes can
  produce physical effects)

- now we can design and build and debug and modify and completely
  explain a class of virtual machines where things are very much
  clearer, including VMs that can do things that previously no
  artificial machine could do, only humans (or other animals,
  in some cases).

What is more, we can use these simple cases to refute incorrect
theories that have confused thinking about the more complex
virtual machines, e.g. theories which claim that every component
of a virtual machine must correspond to a physical part of the
implementing physical machine, and theories that claim that
virtual machine entities and events cannot have physical effects
since they assume that only physical causes can have physical
effects. (A calculation or piece of reasoning based on information
gained through sensors can cause a virtual machine controlling a
chemical plant to send a signal that closes a valve, or increases
the speed of a pump, for instance.)

But more importantly, we have begun to discover some of the huge
variety that is possible in virtual machines enabling us to
explore richer and richer variants, including hybrid virtual
machines with complex architectures that include quite different
sub-machines: for instance machines combining operations on
discrete structures (trees, graphs, lists of words forming
sentences, equations etc.), machines that operate on image arrays
that are treated as samples of a continuous image, and machines
involving highly concurrent networks with continuously varying
levels of activation or connection strengths (as in various kinds
of neural nets).

Thus instead of being confused amateurs studying virtual machines
that are almost completely beyond our comprehension we are
developing professional concepts, formalisms, tools and theories
(including mathematical theories) that enable us to construct ever
more sophisticated virtual machines that we don't merely talk
about but actually build, and demonstrate working, and in the
process discover their limitations and get ideas about further
requirements and further design possibilities for meeting those
requirements.

Unfortunately some people who have studied Turing machines or who
have learnt to write programs in Fortran, or have read an
introductory book on expert systems, think they therefore
understand all there is to know about virtual machines that can be
run on computers.

Nobody understands all there is to know about them: the subject is
still in its infancy (and so-called proofs of their limitations
are based on false assumptions).

Moreover most of the interesting things to be learnt are not just
*general* principles relevant to all virtual machines but what the
*specific* requirements are for virtual machines with specific
human capabilities.

E.g. in designing AARON (no relation) Harold Cohen had to think
very hard, over many years, about how he experiences shapes and
relationships between shapes, and later on how he experiences
colours. (For many years he had to add colours by hand to the
pictures produced by the machine.)

Of course, we can't say that his program has human experiences,
especially aesthetic experiences, but it must have some primitive
grasp of spatial structure and relationships including things like
relations between human body parts, and flowers, tables, doors,
etc. in order to be able to assemble the components of the
pictures it makes.

Likewise some of my colleagues are trying to develop virtual
machines that can find mathematical proofs more successfully than
existing theorem provers, and in order to do that they have to
reflect on their own experiences in searching for proofs, which
requires far more than the shallow knowledge of the low level
syntactic structure of legal proofs.

In particular they are trying to give the machine a kind of
intuitive grasp of the high level structure of proofs so that the
machine can first sketch out abstract proofs (proof plans) and
then try to fill in the details, sometimes even using geometrical
shapes as models, just as human mathematicians sometimes do (e.g.
proving that the sum of the first N odd numbers is a perfect
square, by thinking of ways of arranging dots in square arrays).

I don't know whether AARON works out a sketchy plan for its
picture before working out where the detailed lines go, but that
ability to move up and down between different levels of
abstraction is a feature common to a lot of human thinking.

So replicating that would be part of the task of a machine that
shared qualia with a mathematician: reasoning qualia?

However the notion of qualia that has come out of philosophy and
been uncritically accepted by so many scientists is too
ill-defined for it to be clear what sort of virtual machine would
have qualia. People can't even agree as to whether humans have
them, or house-flies have them, an indication that the notion may
involve conceptual confusion.

I suspect that the deepest theory we'll ever have about qualia will be a
theory that tells us how to build robots with virtual machines so like
those of humans that the robots go through the very same virtual machine
states and processes that led philosophers to invent the notion of
qualia, feels, what it is like to experience something, etc. That's
still a long way off, as the task is so unclear.

However, a possible line of development flows from the observation
that humans can switch attention between all sorts of things.

This might be implemented in a virtual machine that selects
different kinds of information to process and processes it in
different ways: e.g. switching between attending to the door and
attending to the window, and while attending to the door switches
between attending to its shape, its colour, its texture, its
style, etc. (Virtual Machine Attention System: VMAS).

Suppose you built a robot that could not only move around, and see
things using something like TV cameras, but, using VMAS, could
also switch its attention around, sometimes attending to past
events, sometimes thinking about what may happen in the future,
sometimes looking at the shapes of objects, sometimes attending to
their colours or textures. Suppose that it could also turn its
attention *inwards*, for instance noticing what it was thinking
about, or remembering how it had thought about this topic
previously, or examining not the external objects perceived
visually, but also some of the internal intermediate results of
visual processing.

Such a robot might then notice that not only do things in the
environment change, e.g. if they are moved, painted, broken, etc.,
but even when nothing in the environment changes something can
change in its perception of the environment. E.g. where
experienced edges of objects meet, or how acute the angle appears
where two edges of a table seem to meet, or whether one thing
feels warmer than another to the touch.

Some of these are facts about visual appearances that humans don't
notice until they are trained to drawn and paint objects realistically.
The same might be true of a robot: it might go through life never
noticing the non-rectangularity in the experienced shape of a table top
(as a young child apparently doesn't).

But if the robot does manage to switch its 'inner' attention onto those
aspects of how it perceives external objects, it may start thinking:
There is something in me which is the content of my sensory experience
which is distinct from the external objects that stimulate me to have
those contents. The more intelligent robots will short-circuit a few
hundred years of philosophy and reinvent most of the the philosophical
theories about consciousness, qualia, the relation between mind and
body, whether mental events can cause physical events, and whether they
have free will.

However, producing robots that are capable of replicating the internal
states and processes that have led to human theories and debates about
consciousness will require an information processing architecture that
is capable of supporting all those states and processes. But at present
we don't really know what that architecture will be like. Some of us
have first draft speculations regarding the architecture (e.g. the
H-Cogaff architecture described on my web site) and the similar but
slightly different architecture described in Minsky's draft book, The
Emotion machine, on his web site.
    http://web.media.mit.edu/~minsky/

At this stage besides the task of learning how to design sufficiently
complex and realistic virtual machines and show how they might be
implemented in brains, we have the additional and possibly harder task
of analysing the *requirements* for those machines.

It's hard for many reasons including the fact that we don't know
and can't easily find out what's actually going on in human
virtual machines (for the sorts of reasons that have been
mentioned previously, including limited bandwidth for external
communication, and also restrictions on internal self-monitoring
and self reporting).

Moreover, the enormous versatility of human virtual machines means
that almost anyone studying them will miss out on a lot. But
different investigators, especially investigators from different
disciplines can learn different things, and we can pool the
results to begin to build up a list of specifications for the
design of a machine with human-like internal (virtual machine)
processes.

Of course, some people (and some robots) will argue that even when
we have a complete or nearly complete set of specifications and
have built a working system that meets the specifications, not
only in its externally observable behaviour but also in its
internal virtual machine operations, it is still missing something
that humans (or really intelligent robots) have.

When the virtual machines in our robots go through processes of
attending inwardly (using VMAS) and reasoning about the processes,
which lead them to the conclusion that our virtual machine design
could not possibly explain what is going on inside *them* then
we'll have good reason to think our design is at least
approximately right as a theory of how human minds work (until a
better theory comes along).

That will still leave the problem of explaining in detail how the
virtual machine is implemented in brains. That is not the same
thing as finding neural correlates of bits of virtual machines.

In general implementation of virtual machines will be more
holistic. But that's another story.

I'll attach this message to

    http://www.cs.bham.ac.uk/research/cogaff/misc/consciousness-requirements/

Aaron