From Aaron Sloman Tue May 16 11:47:38 BST 1995
To: dave@twinearth.wustl.edu
Subject: more on consciousness
Cc: korb@cs.monash.edu.au

I've now (quickly) read "Facing up to the problem of consciousness")

I thought I would copy my comments to Kevin, to see what sort of
reaction he has -- if he has time to read them.

Your paper is very clear and I agree with a lot of what you say about
what's wrong with other theories. My own position is very similar to
yours, in ways that may surprise you. I think merely that you
over-emphasis physical structure and underestimate the significance of
causal/functional role in determining information content, as I'll
indicate below.

I think you are helping me to clarify my own view about qualia (I still
don't want to say there's any one thing that "consciousness" refers to,
but certainly people who talk about it in theoretical contexts are,
among other things, referring to (sensory) qualia.)

I agree that we need a non-reductive theory of experience, etc. What you
don't point out is that we need non-reductive theories for all sorts of
phenomena which in some sense are implemented in the physical world
(i.e. they are supervenient on physics) but they are not reducable to
physics in the sense that

    (a) the concepts required for describing them are not definable
    in physical terms (at least not by any of the standard methods of
    defining a new concept in terms of existing ones)

    (b) the laws of their behaviour are not reducible to the laws of
    physics.

    (c) there is enormous variability in how any particular supervenient
    system may be implemented in physical systems, so empirical
    correlations are not to be found.
    (I suspect there's no set of physical conditions that are either
    necessary or sufficient for the existence of a typical supervenient
    phenomenon of the sort described below, but I won't argue that
    here.)

Consider an office automation system. I can describe it as acquiring,
storing, using, changing, information about many things, such as the
salaries of employees, unpaid bills, the addresses of customers, how
certain kinds of information are to be processed, what staff are allowed
or not allowed to do, etc. None of the concepts I have used, including
the notion of "information about" can be defined in terms of physics.
(That takes some arguing, but I expect you will agree.)

Moroever, the laws or rules governing the information processing
behaviour of the system are not derivable from physical laws. To see
this consider that they can be changed without changing physical laws.
E.g. we can change the system to prevent managers looking at email of
their staff. We can change the word processor software so that all
documents come out using helvetica font. We can change the time after
which non-payment of bills causes a reminder to be sent, etc. etc.

In the same sort of way there are diverse social, political, economic,
phenomena that are not reducible to physics: the customs or laws about
marriage or the organisation of the economy can change without there
being changes in physical laws.

The physical infrastructure of the universe is capable of supporting
infinitely many different kinds of non-physical systems with different
sets of non-reducible laws of their own.

Of course there will be physical changes underlying the changes in the
"higher level" laws, though not changes in the physical laws. But the
physical changes that support changes in the information processing
rules, changs in the social or political structure etc. will often be
impossible to identify, and moreover, there is no simple, law-like
relationship between the physical changes and the "higher level" changs.

Changing the rules of chess on a sparcstation will involve totally
different physical changes from those involved in changing the rules of
chess on a computer of the 1960s, with magnetic cores, etc.

Moreover, the non-reducible phenomena that I am talking about can enter
into causal relationshipos. An economic depression can cause people to
lose their jobs. Losing a job can cause one to become badly fed, or even
to become ill and die. I.e. the causal relations not only hold between
the non-reducible phenomena: they can even have physical effects.

So my first point is: that much of what you say about experiences, or
qualia, is true, not only of them, but also of many other things.

I think qualia can be causes as well as effects: otherwise their
pervasive existence would be a mystery from an evolutionary point of
view. I guess this is also implied by your comment:

    Briefly put, we can think of awareness as *direct availability
    for global control*.

since control implies causation.

However, when you say:
    Wherever there is conscious experience, there is some corresponding
    information in the cognitive system that is available in the control
    of behavior, and available for verbal report.
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

the last bit is wrong. I see no reason to suppose that the contents are
always available for verbal report. For example, I have no reason to
doubt that monkeys have them, but I have every reason to doubt that in
monkeys they are available for verbal report, or indeed that most of a
monkey's experiences are available for any sort of *external* report.

That's because the main sort of control function of qualia concerns
*internal* behaviour, not external behaviour. Compare: there are aspects
of the functioning of an operating system (e.g. interrupt mechanisms)
that are used to control its internal behaviour but are hard for it to
report: because any processes that could provide the reports would
interfere with the processes themselves, and change them. We may need,
in future, to design new sorts of computing systems to get round this.

(Neural mechanisms, where additional "output" channels can coexist with
the main functional ones in a net may naturally have the advantage in
this respect. Similarly if computers can be developed so that, e.g. one
component can use the electromagnetic radiation of another component
unobtrusively to record and observe what's going on, then they may be
more amenable to external reports of deep internal processes. Richer
internal control operations will then also become possible.)

    It is a central fact about experience that it has a complex
    structure.

Agreed. Bernard Harrison, many years ago, wrote a little book called
something like "Form and Content, in which he argued that such structure
existed even for colour qualia, so that the usual discussions about
inverted spectra were internally inconsistent: the supposed inversion
could not preserve the structural relationships, including links with
non colour concepts (e.g. some colours are warm, others soft, etc.),
which in turn have links with other things.

    Even emotions have structural properties, such as relative
    intensity, that correspond directly to a structural property of
    processing; where there is greater intensity, we find a greater
    effect on later processes.

In many cases there are not two separable things. The inability to take
your mind away from thoughts about the lost child is not caused by the
intensity of the anxiety, it IS the intensity.

Emotion qualia (at least those that involve our high level cognitive
processes, unlike, e.g. a startle or physical disgust) are very
different from sensory qualia. That does not contradict your main point,
however, with which I agree. Roughly speaking you (and I) are
challenging old ideas such as that the soul is indivisible. The soul
could not do its job if it were not a machine with intricate parts.
(Putnam says something similar.)

However, you continually emphasise reportability, and that's a mistake,
for that's a feature of only a small subset of the phenomena. (I think
Dennett made a similar mistake in his early writing about consciousness,
which he wished to link to a speech centre.)


    This principle has its limits.  It allows us to recover structural
    properties of experience from information-processing properties, but
    not all properties of experience are structural properties.

I agree, but not for the reasons that you give immediately after.

In particular the main reason is that there are properties arising
from functional roles which are different from the structural
properties. Consider visual intermediate level databases recording 2-D
structure of the optic array: besides having one lot of functional roles
because of their structure and connectivity to various control
mechanisms (e.g. both posture control and automatically feeding
information to higher level databases concerned with 3-D structure of
the scene) they may have other functional roles because some less
automatic processes can switch attention to the intermediate databases
and inspect their structure (as people have to do in learning to draw,
for example, or in sighting a gun, or in my case reverse-parking a car).

I don't know how many other animals that have visual experiences also
have the freedom to switch attention between levels of processing as we
do (or some of us -- maybe not young children? Maybe not some people
with brain damage?). If other animals can't then their visual experience
must be different from ours. (More on this below.)

    There are properties of experience, such as the intrinsic nature of
    a sensation of red, that cannot be fully captured in a structural
    description.

Yes because of their multiple functional roles, including their roles
within both attentive and non-attentive processing.

    The very intelligibility of inverted spectrum scenarios, where
    experiences of red and green are inverted but all structural
    properties remain the same, ....

I am not convinced that this is intelligible, except in the superficial
way that it's intelligible to say that everything in the universe is
shrinking but all the laws of physics continually adjust so as to make
the shrinkage undetectable.

    ... show that structural properties constrain experience without
    exhausting it.

That follows from the fact that besides (internal) structural properties
there are external structural relations (with other potential
experiences) and functional relations.


    Nevertheless, the very fact that we feel compelled to leave structural
    properties unaltered when we imagine experiences inverted between
    functionally identical systems ....

There's no such compulsion. It's a game philosophers play and they have
freely invented it. I find it not at all tempting, let alone compelling!
In fact, I suspect that most people do not imagine this at all. They
merely imagine that they imagine it (i.e. they describe the imagining to
themselves, but they don't really do the imagining).


    Empirical investigation might even lead us to better understand the
    structure of awareness within a bat, shedding indirect light on
    Nagel's vexing question of what it is like to be a bat.

Here I am on Nagel's side. You may acquire more and more information
about the structure and function of a bat's experience, but because you
are a very different information-processing system with quite different
semantic structures and semantic variability there is no way your brain
can be put into a state that amounted to your "knowing what it is like
to be a bat".

(I think this is connected with Wittgenstein's remark that if a lion
could talk we could not understand him, but I am not sure that he meant
this).

My point (which others have made) is that the semantic capabilities of
two information processing sytems may be sufficiently different that
there is NO way of translating from one to the other. There may be
partial communication of information, but that's all.

In that sense, no adult can understand what it is like to be a young
child, and we cannot even remember what it was like: though we can fool
ourselves that we do. We process records in terms of our present
conceptual apparatus, which was not available during the original
experiences.

Similarly much of the time when we talk about what our pets think, want,
feel, etc. we are inevitably anthropomorphising. I am not saying that
they have nothing like what we attribute to them, just that the
attributions are often vastly over-specific.


    2. **The principle of organizational invariance**.  This principle
    states that any two systems with the same fine-grained *functional
    organization* will have qualitatively identical experiences.

Yes except for the fact that much of our way of talking about
experiences implicitly or explicitly defines them in terms of their
relations to some external object. For instance, if I describe Fred as
thinking about visiting the Eiffel tower, then I am referring to an
object outside Fred in describing Fred's mental state. Another Fred
in another part of the universe might have exactly the same internal
states and processes yet be thinking about another tower.

That's a trivial example. I suspect there may be far more subtle ways in
which we implicitly, and unwittingly refer to the environment when we
describe internal states. (Maybe even when we talk about the experience
of seeing red.) In that case the reproduction of the internal structures
will not suffice for reproduction of those states, so described. (I
don't expect you to disagree with that: it's just a messy complication
that needs to be got out of the way.)


    It is a central fact about experience, very familiar from our own
    case, that whenever experiences change significantly and we are
    paying attention, we can notice the change; if this were not to be
    the case, we would be led to the skeptical possibility that our
    experiences are dancing before our eyes all the time.

Dennett reports experiments in which visual displays change during
saccades: the "subject" fails to see the changes, whereas others
watching find the "dancing" totally obvious. If true, this refutes the
most obvious interpretation of what you are saying, though the example
you were discussing was very different in structure.

I suspect you are (like many people) confusing empirical claims about
what we can or cannot do, with logical claims about what it means to say
or think "this is how it appears to me", on which (by definition) one
cannot be mistaken (though one can be insincere, etc.).

    To put it in technical terms, the philosophical hypotheses of
    "absent qualia" and "inverted qualia", while logically possible, are
    empirically and nomologically impossible.

This seems to contradict what you earlier said was compelling.

It is not at all clear to me that you have produced any argument for the
impossibility apart from the sort of argument that leads some
philosophers to say that those things only appear to be logically
possible but are not really. The notion of "(im)possibility" that we use
in these discussions still needs a lot of analysis. I think Dennett's
attack on this stuff is quite good.


    The double-aspect principle stems from the observation that there is a
    direct isomorphism between certain physically embodied information
    spaces and certain *phenomenal* (or experiential) information
    spaces.

I agree with the sort of thing you are trying to do here, but I think
the details are wrong, especially your reliance on *physical* structure
and isomorphism.

Often the relevant structure exists only in a *virtual* machine. E.g.
consider the 8 bits used to represent a character in many programming
language implementations. Those 8 bits might be implemented in the
physical machine as a 10 or even a 16 bit structure in order to support
error detection and correction at a low level. Or they might be
implemented in fewer bits where data-compression mechanisms are used.

Either way there are *really* exactly 8 bits at the level of the
relevant virtual machine (e.g. there are virtual machine instructions
for separately testing or changing the states of exactly 8 different
bits), and that's independent of the exact physical structure supporting
it. (Sparse arrays are another favourite example.)

I think Newell and Simon make the same sort of mistake in connection
with their phrase "physical symbol system". Most of the relevant symbols
are not physical, but data-structures in virtual machines.

A more subtle and important disagreement I have with you (and Shannon)
concerns your attempt to pin information directly onto structure. That's
a purely syntactic notion of information (a different version is to be
found in algorithmic inforation theory). There's another notion of
information that depends partly on structure, partly in functional
roles.

E.g. bit-pattern B1 may implement a machine instruction, bit-pattern B2
a text character in a word processor, bit-pattern B3 a numerical index.
All three of B1, B2, B3 may have identical bit structures. Their
different "meanings" arise from the different ways in which they are
(created and) used by the rest of the system, which, among other things,
also determines what that pattern excludes. E.g. there may be only 100
different machine instructions, 256 different characters, and an
indefinitely large number of (variable bit-length) integers. Thus, B1
excludes only 99 options, B2 257 options, but B3 indefinitely many
options.

Thus what they mean to the system depends not only on their functional
role but also what the kind of variability there is within that
functional role. (Shades of Saussure?)

So I think this bit needs to be withdrawn or rewritten:

    we can note that the differences between phenomenal states have a
    structure that corresponds directly to the differences embedded in
                               ^^^^^^^^
    physical processes; in particular, to those differences that make a
    difference down certain causal pathways implicated in global
    availability and control.  That is, we can find the *same* abstract
    information space embedded in physical processing and in conscious
    experience.

Change "physical process(es)/(ing)" to "virtual machine process(es)/(ing)"
and it's much better. The same goes for several other passages in your
paper. But further changes are need to accommodate the importance of
role, as opposed to structure, in defining experience.

This is all part of the reason why, although we may be able to describe
(and explain) in considerable detail what it is like to be a bat we
cannot "know" or "imagine" what it is like to be a bat. The bat's
experiences are part of a system within which there are functional roles
that simply cannot be replicated within ours, and vice versa. (Similar
arguments point to fundamental limitations in human communication, but
that's another whole story.)


    Of course, the double-aspect principle is extremely speculative and
    is also underdetermined, leaving a number of key questions
    unanswered.  An obvious question is whether *all* information has a
    phenomenal aspect.

Notice that if you shift the emphasis from information in structure to
information in structure+role then there is always a phenomenal aspect
to information, in the sense that we can talk about what that structure
means to that system, and that will be something that another system
cannot necessarily share or "directly experience" for itself.

I.e. there's something that is inherently subjective about information
content, though not in any mystical or anti-scientific way.


    Where there is simple information processing, there is simple
    experience, and where there is complex information processing, there
    is complex experience.  A mouse has a simpler information-processing
    structure than a human, and has correspondingly simpler experience;
    perhaps a thermostat, a maximally simple information processing
    structure, might have maximally simple experience?

Yes. (I've written very similar things, including comparing the
thermostat as a limiting case to a circle as a limiting case of an
ellipse, where many of the interesting properties of most ellipses are
lost.)

    For example, it is often noted that physics characterizes its basic
    entities only *extrinsically*, in terms of their relations to other
    entities, which are themselves characterized extrinsically, and so
    on.  The intrinsic nature of physical entities is left aside.

 From what I recall of reading Leibniz he took the position that you
are speculating about. Every atom experiences the rest of the universe
in its own way. (This is even more primitive than the thermostat, if
that "experience" is not part of a control mechanism. But perhaps it
is!)

    This metaphysical speculation is probably best ignored for the
    purposes of developing a scientific theory, but in addressing some
    philosophical issues it is quite suggestive.

No, it HAS TO become part of science, for understanding it is essential
to the scientific explanation of the human tendency among people to want
to reify consciousness and say that it is irreducible, etc. etc.

But it won't be part of physics, any more than computer science (or the
science of software systems) is part of physics. It could be part of a
new expanded truly scientific psychology, replacing the blind
correlationism of much modern psychology.

Incidentally, some of the points I've been making are in my Royal
Society paper, Semantics in an intelligent control system, available as

    ftp://ftp.cs.bham.ac.uk/pub/dist/cog_affect/Aaron.Sloman_semantics.ps.Z

and less clearly in an older paper The Mind as a Control System in a
book edited by Hookway and Peterson, available in the same directory as

    Aaron.Sloman_Mind.as.controlsystem.ps.Z

If you disagree with what I am saying and want to attack it publicly you
can refer to those!

Must go now. I hope this makes some sense. I think we are on the same
side, both against the Searles and Penroses, and the mystics, and the
brain scientists who look in the wrong places, and the Dennetts who say
there's no problem.

Aaron