Counterfactual
computational
vehicles of
consciousness
Ron Chrisley
COGS/Dept. of Informatics
University of Sussex
Toward a Science of Consciousness, Tucson
April 7th 2006
Outline
• Bishop's argument against computational
explanations of consciousness
• My response: Acknowledge the counterfactual
nature of physical states
• Segue: Use emphasis on counterfactual
properties as motivation for a specific form of
computationalism/representationalism
• Provides a plausible yet non-trivial enactivist
model of perceptual experience: ImaginationBased Architecture
Bishop: "Dancing with Pixies"
• Poses a dilemma for
computational explanations of
consciousness
• Horns based on two notions of
computation:
1.Non- or weakly- causal construal of
computation
2.Strong, counterfactual causal
construal of computation
Bishop's dilemma
Either notion has problems:
• Horn 1: Weak causality:
– Implies every computation is
realised in every physical system
– So any claim that a given
computation is sufficient for
consciousness implies panpsychism
– Phenomenal "pixies" everywhere!
• Horn 2: Strong causality:
– Violates naturalism by appealing to
non-physical aspects of a state
Rejecting the first horn
• Yes, weakly causal construal of
computation implies
panpsychism
• But:
– What's wrong with panpsychism
anyway?
• Actually, a lot…
– Better (cf Chalmers 94, 96;
Chrisley 94):
• Weak construal not really a causal
construal of computation at all
• Thus does not capture what is meant
by computation
Embracing the second horn
• Strong, counterfactual causal construal
of computation
– Identity of a computational state depends
not only on actual causal relations…
– …but also on causal effects (output,
successor state) a state would have had
were different input received
• Bishop: Subject to variants of
Chalmers' Fading Qualia and Suddenly
Disappearing Qualia arguments
Bishop's thought experiment
• Consider the operation of two
robots:
– R1: "controlled by a program
replicating the fine-grained functional
organisation of a system known to
have phenomenal states"
• A particular run of R1 with input I results
in an actual sequence of behaviours B
– R2: any open physical system that
generates B, given the same input I
More on R1 and R2
• For example, R1 might be
controlled by an AI program that
enables it to output classifications
of objects presented to its
cameras
– When given the input of a particular
object, this results in a particular
sequence of output classifications B:
– "This colour of this triangle is a bit
more red than the square I just saw"
• While R2 can just have a hardwired circuit that happens to
output B (regardless of input!)
Branching FSA
• We can conceive of R1 and R2 as a
finite-state automata with branching
and non-branching states, respectively:
Non-branching FSA
(Diagrams from Bishop 2002)
The computationalist's view
• Bishop: "Hence, although the external
behaviour of the two systems over the
time interval is identical [viz, B], for [a
computational theory of
consciousness], only R1 would
experience genuine phenomenal
states."
• What's wrong with that?
Transforming R1 into R2
• One by one, delete one of the N state
transition sequences of R1 that are not
actually used in the case under
consideration, to transform R1 into R11,
R11 to R12… to R1N
• R1N will be computationally formally
identical with R2
• So for a computationalist, R1N, like R2,
has no conscious experience
R1, R2 & another dilemma
Bishop:
• "What happens to the phenomenological
experience of R1 as it incrementally
undergoes the above transformation?"
• "Either its experience of phenomenal states
must gradually fade (Fading Qualia) or it must
switch abruptly at some point (Suddenly
Disappearing Qualia)."
Me:
• Not necessarily: There might be several,
spaced, discrete transitions.
• But let that pass…
No SDQ?
Bishop:
• Rule out first horn: Suddenly Disappearing
Qualia
• It would "imply that the removal of one such
privileged branching state transition
instruction would result in the complete loss
of the robot’s phenomenal experience"
Me:
• Not convinced this is a problem
• But agree to rule it out for the sake of
argument
A general argument?
• Bishop presents an argument not against the
second horn (Fading Qualia), but against
computationalism in general:
• The computationalist's position implies the
existence of "a system, whose phenomenal
experience is contingent upon non-physical
interactions with sections of its control
program that are not executed – a form of
dualism."
• "Hence, if phenomenal states are purely
physical phenomena, the phenomenal
experience of the two robot systems, R1 and
R2, must be the same."
Warning sign: Too strong
• An indication that Bishop's
argument can't be right:
– It proves too much
• If right, it would imply that there
could never be a physicalist
computational explanation of
anything…
– …not even computers!
Misunderstanding the physical
• Bishop's main mistake: claiming that
differences in counterfactual behaviour
do not constitute physical differences
• Presumably, it is by virtue of some
physical difference between a state of
R1n and the corresponding state of
R1n+1 that gives the former a
counterfactual property the latter lacks
Misunderstanding the physical
• Note that to delete the nth transition,
one would have to physically alter R1n-1
• So despite Bishop's claim, if R1 and R2
differ in their counterfactual formal
properties, they must differ in their
physical properties
• Causal properties (even counterfactual
ones) supervene on physical properties
Counterfactuals are key
• So much for Bishop's argument against
computational accounts of
consciousness
• But although Bishop has nothing on
computationalism in theory, he inspires
a relevant critique of the form it usually
takes
• That is, standard computationalist
theories of consciousness neglect the
importance of counterfactual properties
Segue…
"Actualist" computationalism
• Typically, computationalist (or
functionalist) theories attempt to
map:
– A perceptual phenomenal content
– To a computational (functional) state
– By virtue of the latter's actual causal
origins (and perhaps its actual causal
effects)
The Grand Illusion?
• For example, some argue:
– Change blindness data show
that only foveal information
has an effect on our
perceptual state
– Thus, our perceptual
experience is only of the
foveated world
– Any appearance that anything
else is experienced is
incorrect
QuickTime™ and a
TIFF (Uncompressed) decompressor
are needed to see this picture.
Being counterfactual
• But a computationalist theory that places explicit
emphasis on the role of counterfactual states can
avoid the Grand Illusion result
• E.g.: The phenomenological state corresponding
to a given computational state includes not just
current foveal input
• But also the foveal input the computational
system would expect to have if it were to engage
in certain kinds of movement
• "Imagination-based architecture" (IBA)
More on IBA
• These expectations can be realized
in, e.g., a forward model, such as
a feed-forward neural network
• The model is updated only in
response to foveal information
– E.g., it learns: "If I were to move my
eyes back there, I would see that
(the current foveal content)"
The IBA explanation
• Thus, change blindness can be
explained without denying peripheral
experience
• Consider the system after an element
of the scene has changed, but before
the system foveates on that part of the
scene
• The expectations of the forward model
for what would be seen if one were to,
say, foveate on that area, have not
been updated
No Grand Illusion
• According to IBA, the
(outdated) expectation
is a part of current
experience
• Thus no change is
detected or experienced
• So our experience is
just what it seems
QuickTime™ and a
TIFF (Uncompressed) decompressor
are needed to see this picture.
Elaborations to IBA
• Only a simplistic version of IBA
presented here
• Can be elaborated to include
change not instigated by the
system itself
– E.g., expectations of what foveal
information one would receive if the
world were to change in a particular
way
More elaborations to IBA
• Weighted contributions to
experience
– Current foveal info strongest of all
– Expected foveal after a simple
movement a little less
– Contribution of expected results of
complex movements/sequences
inversely proportional to their
complexity
Open questions for IBA
•
E.g., what is the experience at a nonfoveated part of the visual field if one has
different expectations for what one would
see depending on the motor "route" one
takes to foveate there?
–
–
–
–
Some "average" of the different expectations?
Winner take all?
Necker-like shift between top n winners?
No experience at that part of field at all, as
coherence (systematicity, agreement) at a time is
a requirement for perceptual experience?
Announcements
• For philosophers of Cognitive
Science/AI:
– My department, Informatics at the
University of Sussex/COGS is
hiring
– Tell your friends/colleagues!
Quick Time™ and a
TIFF (Uncompressed) dec ompressor
are needed to s ee this pic ture.
• Those interested in Machine
Consciouness:
– Conference I am chairing, BICS
2006 in the Greek Islands, October,
is still accepting submissions to the
end of April
• Email me about either/both:
[email protected]
BICS
Thank you!
• Thanks to Mark Bishop and Rob Clowes
for helpful discussions