[Converted to html 4 Jun 2013]
PDF Version: http://www.cs.bham.ac.uk/research/projects/cogaff/81-95.html#52
Commentary on
Max Velmans, `Is Human Information Processing Conscious'
in
Behavioural and Brain Sciences
C.U.P. 14-4,pp694-695, 1991
DEVELOPING CONCEPTS OF CONSCIOUSNESS
Aaron Sloman
Velmans cites experiments undermining hypotheses about causal roles
for consciousness in perception, learning, decision making, and so on.
I'll leave it to experts to challenge the data, as I want to concentrate
on removing the surprising sting in the tail of the argument.
Having argued that consciousness has no functional role, he does
not conclude that it is just a myth (like the aether). Instead, he
argues that because there is no functional "third-person" status, it
must be given some other kind of "first-person" analysis. He also says
that `Consciousness is, nevertheless, amenable to scientific
investigation', and that a complete psychology requires two
complementary, mutually irreducible, perspectives, one of which studies
consciousness.
Curiously, despite discussing subjects' reports on conscious
episodes, he does not consider the possibility that the functional role
might include enabling individuals to give others information about
their mental states (useful in many contexts: in the family, in the
classroom, in the dentist's chair, etc.)
I won't defend this view because, as explained below, the word
"consciousness" is associated with too many muddled ideas for any such
statement to be worth defending. Instead, I'll propose a different
approach, from an engineering standpoint (Dennett's `design stance',
Dennett 1978). This views humans (and other animals) as having a
complex and sophisticated design (which is not to say that there's any
designer), and attempts to consider how one might make something with
similar capabilities. Design provision must be made for "consciousness",
"awareness" etc. (See also Dennett 1983)
Unfortunately it is not at all clear what this means. Because
"consciousness" is ill-defined, Velmans says `...it is consciousness in
the sense of "awareness" that is of primary concern'. Is "awareness"
any clearer? Is the fly aware of my approaching hand? Is a dreamer
aware of the pain, or the pursuing lion? Most people say: "Yes, that's
why dreams are nice, or frightening." Others might think you can't be
aware (= conscious) when you are asleep (= unconscious). Is the
sleepwalker aware of the door-handle when he looks at it and turns it?
We are all aware that the end of the century is approaching. If this is
included in `the sense of "awareness"' then consciousness includes
nearly everything we know. Where are the boundaries?
Velmans is apparently mainly concerned with self awareness, i.e.
inwardly perceiving one's internal states and processes. (Perceiving
-1-
them, not just knowing about them?) There are many situations in which
we are self-conscious or self-aware in some way. But not throughout
waking life: when you are totally absorbed in something outside you,
e.g. watching an exciting football match, or a gripping play, do you
then lack "first-person" consciousness?
Conjecture: people who discuss consciousness delude themselves in
thinking that they know what they are talking about. I don't claim that
there is nothing they are talking about. Rather, it is not just one
thing, but many different things muddled together. That colloquial
language uses one noun is no more evidence for a unique reference than
the multifarious uses of the word "energy" (intellectual energy, music
with energy, high energy explosion, etc.)
Why not, like physicists, ignore colloquial usage and agree on some
technical definition of the word "consciousness"? Partly because there
is also considerable emotional energy associated with the word, which
will interfere with serious usage of any technical homonym. More
important, we don't yet have an adequate understanding of the issues: We
don't know what the relevant capabilities of human beings and other
animals are; we don't know what functional decomposition (i.e. what
sort of architecture) underlies these abilities, and we don't know what
sorts of mechanisms (electrical, chemical, neuronal, software, or
whatever) are capable of producing such functionality.
Claiming to know what consciousness is by attending to it is no
more convincing than claiming to know what spatial locations are by
attending to them. It didn't help Newton. Attending doesn't answer
questions about identity: "When is another thing the same place (or
mental state) as the one attended to?" has different answers depending
on what relationships are in question. A fly, a mouse and a person may
all be aware of a moving object: Is that the same state?
There's no answer because there's nothing unique that you have and
others definitely do or do not have. I am not denying the existence of
what's attended to -- just its unique identification. Your state is very
complex and other things may have states that are partly similar, partly
different. But in what ways? How many different substates underlie
"conscious" states? What feels like something simple that is either
present or absent, is actually something complex, aspects of which may
be present or absent in different combinations. (This is not a matter of
degrees, but of kinds of different states.)
If we give up the idea of a unique referent, we can instead survey
relevant phenomena, analyse their relationships to other capabilities,
and then attempt to come up with explanatory designs: a hard task.
Psychologists, philosophers, linguists, anthropologists, hypnotists,
artists, and others can explore, survey, and summarise the many and
varied phenomena. Cognitive designers can work both bottom-up trying to
extend existing computational models and top-down, trying to produce
detailed requirements analyses and design specifications for systems
with human-like capabilities. Then, after analysing the design-
tradeoffs, we can try devising mechanisms capable of generating all
these capabilities, including self-monitoring capabilities.
An architecture that supports not only the perception of external
events but also the monitoring of relatively global and abstract
internal states could have a number of features consistent with the data
reported by Velmans.
If sensory mechanisms monitoring the environment can themselves be
-2-
monitored, there will be a delay between the occurrence of the first-
order perceptual processes and the results of second-order monitoring,
just as there is a delay between retinal events and the production of
information about the environment. Moreover, self-perception, like all
perception, will involve errors, loss of information, distortion, and so
on.
Similarly, some designs support the ability to monitor "control
processes" that take decisions, form plans, initiate actions. Again one
can expect delay between the occurrence of the first-order internal
events and the production of high level summary information about those
events, and perhaps also some distortions and errors.
It is not surprising that results of monitoring that occur after
the events monitored cannot be causally involved in their production,
but they can still have a causal role: informing others, keeping records
of internal processes for many purposes, including long-term feedback
that revises strategies used in first-order processes. Experiments
showing that some kinds of learning occur without high level self-
monitoring do not imply that all do.
Another second-order process is high level decision making. The
various subsystems that produce or control actions of various sorts may
themselves be subject to "meta-level" control. Many parts of the system
will normally chug along on the basis of information available to them.
But some mechanism is required for coping with conflicting needs and for
high-level long-term coordination and strategy formation.
This might use a "democratic" voting system with numerical
summation and comparison procedures (e.g. in neural nets).
Alternatively, high-level strategy formation and conflict resolution
might be reserved for a special subsystem with access to more
information and more powerful reasoning capabilities than the others
(Sloman 1978). Mixed modes of global control are also possible.
Training could also use second-order processes. Many skills require
low-level mechanisms to be trained by being taken through various steps
in a complex process, with fine-tuning based on feedback (e.g. learning
to drive a car, play the violin, pronounce words). This might use a
mechanism that does partial analyses of the steps required, then guides
the lower levels through those steps, and increasingly lets them take
control. Internal monitoring of the behaviour of second (or higher)
order control facilities would require yet another mechanism, alongside
those monitoring perceptual processing.
These are but hints at the many and varied ways different levels of
monitoring and control may coexist in intelligent agents: some
mechanisms controlling others, some monitoring others, some training
others, some resolving conflicts between others. Perhaps this somehow
produces the illusion that there is one high level process in charge of
and monitoring everything.
Conjecture: This (very difficult) design-based strategy for
explaining phenomena that would support talk of consciousness will
eventually explain it all. We shall have evidence of success if
intelligent machines of the future reject our explanations of how they
work, saying it leaves out something terribly important, something that
-3-
can only be described from the first-machine point of view.
References
D.C. Dennett, Brainstorms Bradford Books and Harvester Press, 1978.
D.C. Dennett, `Intentional systems in Cognitive Ethology', Behavioral
and Brain Sciences, 6(3), 1983.
A. Sloman The Computer Revolution in Philosophy: Philosophy Science and
Models of Mind, Harvester Press, and Humanities Press, 1978.
-4-