Consciousness as a target for the artificial sciences

1
Consciousness as a target for the artificial
sciences

• Steve Torrance
• March 2007

2
Consciousness as a target for the artificial
sciences.

• The topic of this talk is The realizability of
  Artificial Consciousness. The basic message is
  Its a long way off. No surprise there, perhaps.
  Except that some tenaciously beg to differ and
  see it as imminent.
• The point of the discussion is that dealing with
  the question in proper depth helps us to get a
  better grip both on the nature of (actual and
  possible) consciousness and on the scope and
  limits of the sciences of the artificial.
• Ill make three distinctions. First, between
  strong and weak artificial consciousness
  (AC). Its mainly the strong AC project that
  Ill be considering here. (Actually AC
  researchers are rather coy about whether theyre
  pursuing the first, the second, or both.)
• Second, between thin and thick conceptions of
  consciousness. Ill argue that an inadequate,
  thin conception of consciousness dominates the
  thought of many devotees of AC, as well as the
  thought of defenders of traditional
  (anti-physicalist or anti-computational)
  approaches to consciousness and mind.

• Third, theres the distinction between
  functional and phenomenal consciousness.
  Many AC workers see the former as a kind of
  half-way house to realizing strong AC. Ill
  argue that you cant have functional
  consciousness alone. When we talk about
  functional consciousness were talking about the
  functional aspects of a (phenomenally) conscious
  organism.
• On the thick conception of consciousness, there
  are perhaps different ways to elaborate this.
  Ill sketch one approach. This consists in
  seeing consciousness in terms of a kind of Grand
  Inventory of miscellaneous strands. All or most
  of these have to be addressed head-on in any
  serious AC research plan.
• This gives mixed news. Strong AC is a long way
  further off (and may require as yet undreamt-of
  technological platforms). But at least we will
  have a more realistic set of success-criteria for
  the enterprise. This probably has general
  implications for how we view the past and future
  of the sciences of the artificial. And we will
  also have learned more about the nature of
  natural consciousness.

3
Background

• Background to this talk
• Workshop on ethics and artificial agents,
  Baden-Baden, 2004
• AISB Workshops on Machine Consciousness,
  Hertfordshire, 2005 Bristol, 2006
• Forthcoming special issue on Machine
  Consciousness (edited with Ron Chrisley and Rob
  Clowes), Journal of Consciousness Studies, July
  2007.
• The topic the realizability of Artificial
  Consciousness
• The message its a long way off
• The point dealing with the question in proper
  depth helps us to get a better grasp on
• the nature of (actual and possible) consciousness
• the scope and limits of the sciences of the
  artificial

4
Artificial Consciousness as a project

• My concern in this talk is about THE PROJECT OF
  DEVELOPING ARTIFICIAL CONSCIOUSNESS (AC)
• A central question is
• What features of the natural phenomenon of
  consciousness need to be taken seriously by
  workers in the sciences of the artificial
  working on building (AC) systems?

• I propose to draw up a Grand Inventory of such
  features, as a guide for those intent on creating
  AC.
• This is dependent on a certain conception of
  consciousness that I call the thick (or deep)
  conception (as opposed to the thin (or shallow)
  one)
• Much of the discussion will be on the necessary
  preliminary philosophical issues

5
Three distinctions

• Weak versus strong Artificial Consciousness (AC)
• Thin versus thick conceptions of C and AC
• Functional versus phenomenal consciousness

6
Two kinds of research into Artificial
Consciousness

• WEAK AC
• Corresponds to weak AI weak ALife
• Trying to produce systems that clarify aspects of
  biological consciousness,
• but without any pretension to be producing actual
  consciousness in artificial form
• Ontologically innocent
• i.e. doesnt make claims about the conditions for
  the existence of consciousness in artificial
  systems

• STRONG AC
• Cf strong AI strong ALife
• Aiming to generate actual (psychologically real)
  consciousness artificially
• At the moment only computationally-based
  platforms are used
• one day perhaps artificial biology /
  neuro-engineering?
• Ontologically loaded
• i.e. it DOES make claims about conditions for the
  existence of consciousness in artificial systems

7
The distinction between thin and thick
conceptions of consciousness

• Alongside weak and strong AC it is possible
  to put another distinction
• Between thin (shallow) conceptions of
  consciousness
• And thick (deep) conceptions
• This is a distinction between conceptions of
  consciousness per se, not (just) artificial forms
  of C.
• Ill claim that most current approaches to strong
  AC (and maybe weak AC too) rest on a thin
  conception
• and that thin conceptions do not provide an
  adequate understanding of what consciousness is
• To do this, we need to delve into a little of the
  background of standard debates on consciousness
  by certain philosophers (Frank Jackson, David
  Chalmers, etc.)

8
Computationalism, Physicalism and the
anti-physicalist thought-experiments.

• Scientists like to assume that the physical world
  (including brains, bodies, etc.) IS ALL THERE IS.
• So consciousness must have a FULLY PHYSICAL
  EXPLANATION it must be explained totally in
  terms of physical processes in the brain body,
  and their interactions with the physical world.
  (Physicalism)
• Computer scientists often assume, in addition,
  that the brain is a kind of (complex) computer.
  and that consciousness can be explained in terms
  of computational processes realized in the
  physical brain. (Computationalism)

• The strong AI and AC programs are based on the
  assumptions that physicalism and computationalism
  are both true.
• (Maybe weak AI/AC too)
• BUT There are a variety of arguments which aim
  to disprove physicalism
• Such arguments are often based on
  thought-experiments descriptions of scenarios
  that are empirically impossible but conceptually
  possible.

9
Two anti-physicalist thought-experients

• TE1 Color-blind brain science.
• (Jackson)
• A person can know all there is to know about the
  physical explanation of color experience (e.g. in
  terms of brain processes)
• and yet have never seen colors or had color
  experience (e.g. by being congenitally
  color-blind).
• So (its argued) theres something else to
  experience over and above the physical features.

• TE2 Zombie-worlds. (Chalmers)
• If the physical features of brains, etc. fully
  explained experience,
• then it wouldnt be even conceivable that you
  could have a world physically identical to ours
  but without any experience.
• But (its argued) such a world (A Zombie World)
  is at least conceptually conceivable.
• So consciousness must be something ontologically
  over and above the physical

10
Response to the anti-physicalist
thought-experiments

• I wish to propose that
• All such arguments and thought experiments depend
  on a certain conception of what consciousness is
  the thin (or shallow) conception
• There is an alternative, more adequate conception
  the thick (or deep) conception.
• The thin conception rightly emphasizes the
  phenomenal feel of consciousness.
• But it does so at the expense of any other aspect
  of consciousness

• On the thin conception,
• the notion of conscious or phenomenal feel is
  conceptually divorcible from any other features
  in an agent,
• including all cognitive processes, behaviour,
  brain functioning, body-system activity, etc.
• The thin conception tends to lead to a dualist or
  epiphenomenalist view
• On the thick conception,
• consciousness is seen in terms of the real,
  physical characteristics of embodied beings who
  experience conscious subjectivity,
• as well as in terms of the subjective feeling
  itself.

11
More on the thick and thin conceptions

• On the thick conception,
• consciousness is, CONCEPTUALLY, a complex,
  multidimensional natural, biological, physically
  embodied, process
• of which phenomenal feel is just one part (but an
  important part)
• The anti-physicalist thought- experiments gain
  their persuasive force from the thin conception
• The thought-experiments seem harder to launch
  successfully on the thick conception
• -- So the threat of a dualist or
  anti-physicalist conclusion is avoided

• A world of zombies which are physically like us
  but which lack phenomenal consciousness
• seems a confusing idea if consciousness is an
  essentially embodied and biological process
• The idea of a color-scientist who knows all the
  physical facts about consciousness but is unaware
  of what its like to see red
• seems puzzling if the experience of seeing red
  is CONCEPTUALLY bound up with physiological
  processes that underpin that experience

12
Functional vs phenomenal consciousness

• Cf Two Conceptions of Machine Phenomenality
  forthcoming in special issue of Journal of
  Consciousness Studies
• (In this paper I apply the thin/thick distinction
  directly to the domain of Machine Consciousness)
• How does the thin/thick division relate to
  debates over artificial or machine C?
• In order to see this we need to discuss a third
  distinction between phenomenal and
  functional consciousness

13
Artificial consciousness researchers tend to
cognitivize or functionalize consciousness

• When AC researchers talk about consciousness,
  they tend to talk in terms of functional or
  cognitive or access consciousness, rather
  than phenomenal or qualitative consciousness.
• This idea refers to the cognitive roles that
  consciousness appears to perform in our mental
  lives.
• E.g. according to Bernard Baars Global
  Workspace (GW) theory, consciousnesss role is
  explained in terms of its contribution to the
  functional architecture of our cognitive
  processing

• GW Theory
• Consciousness works as a central message-board
  which allows information to be passed to the
  various non-conscious functional agencies which
  are operating in parallel
• Thus According to GW theory, C subserves the
  performance of non-routine tasks, such as
• driving round a narrow mountain pass
• working out a difficult mathematical calculation
• using imaginative rehearsal to plan a motor
  action.

14
The uses of cognitive conceptions of consciousness

• Its fine to be talking about cognitive or
  functional aspects of consciousness if one has no
  pretensions to be doing strong AC i.e. to be
  building a truly conscious machine, system or
  artificial creature
• But often people talk as if they are really
  interested in developing a machine which
• (a) really is conscious, but
• (b) is only functionally conscious

• Such people often then go on to state one of two
  things
• one day, through further developments along
  similar lines, phenomenal consciousness may
  possibly develop because all that phenomenal
  consciousness can be, is very complex and subtle
  functionalities OR
• phenomenal consciousness is just a myth or a
  muddle

15
How AC researchers deal with phenomenal
consciousness

• AC researchers tend to feel awkward about
  phenomenal consciousness.
• Getting too fixated on the phenomenal feel of
  consciousness seems to make AC a doomed project.
• (Its much easier to see how computers can
  replicate the cognitive, intelligent aspects of
  mind, than the feeling, phenomenal aspects.)
• So
• AC researchers adopt various strategies to cope
  with the fact that phenomenal feel is hard to fit
  into a computational framework.

• Here are some
• The eliminativist strategy
• Qualia, phenomenal C, etc are illusory,
  confused notions..
• 2. The functionalist strategy
• Phenomenal C can be subsumed under functional or
  cognitive C.
• 3. The agnostic strategy
• Well never know enough about phenomenal C so
  lets just concentrate on functional C
• 4. The special pleading strategy
• Artificial, computationally-based C would be
  special kind of consciousness, so not directly
  comparable to biological C.

16
So Computationalists also trade on the thin
conception

• All of these strategies tend in different ways to
  downgrade or marginalise phenomenal consciousness
• PLUS
• The computationalist strategies also tend to
  trade on the thin conception of C.
• Eliminativists rightly question phenomenality
  thinly conceived
• Cognitivists are right to believe that there is a
  close connection between phenomenal feel and
  cognitive or functional features (such as global
  workspace, etc)
• And computationalists are rightly impatient with
  the arguments offered against physicalism.

• BUT computationalists are operating with their
  own version of the thin conception
• They take the thin conception for granted
• by assuming that it is the only way phenomenal
  feel can be understood
• And in fact operating with the thin conception
  allows phenomenality to be more easily reduced
  away or eliminated
• So in a certain way opponents of physicalism and
  their computationalist critics feed out of each
  others pockets.

17
Illustration Stan Franklins IDA

• Stan Franklins IDA program is an expert system
  running on a (desktop) computer,
• - which assigns postings to US Navy personnel
• Franklin designed it using Baars Global
  Workspace theory as a model
• Franklin claims that it has functional, but not
  phenomenal, consciousness.

• Many other AC models use GW theory, or other
  cognitive/functional models of consciousness
• BUT
• Is functional consciousness, consciousness?

18
The billion dollar challenge (Bringsjord)

• Selmer Bringsjord (JCS, forthcoming) imagines
  being offered 1bn to develop a phenomenally
  conscious artificial being
• Are there any circumstances in which it would be
  prudent, or indeed morally acceptable, to accept
  the challenge?
• This is clearly a strong AC project
• Producing merely functional consciousness
  wouldnt suffice to gain the 1bn.

• But could one produce a functionally conscious
  artifical being (which one wasnt claiming to be
  phenomenally conscious) as a kind of half-way
  house to strong AC?
• (1 bn challenge?)

19
The idea of functional consciousness as a
half-way house

• Could you be claiming to have produced real
  consciousness if all you could claim to have
  produced was functional consciousness
• That is if all you had was a device that
  produced all the cognitive aspects of
  consciousness without the phenomenal feel?

• Argument against this
• Functional consciousness is not a separate form
  of consciousness which is an optional extra to
  full phenomenal consciousness, but
• the functional aspects of the integral phenomenon
  of consciousness
• i.e. the functionality of a conscious being

20

• If you have simply reproduced the functionality,
  you have a non-conscious system which functions
  in the way a conscious system would thats all
• That is its NOT the case that
• (a) there are two different ways in which a
  system can be conscious
• functional phenomenal (or both)
• NOR is it the case that
• (b) You can have a system that is conscious in
  functional way independently of its being
  conscious in the phenomenal way.
• So the idea of targeting functional
  consciousness as a half-way house to strong AC
  doesnt work
• Not even worth 10 cents

21
(No Transcript)
22
A wrong way to put the objection to the half-way
house proposal

• HOWEVER
• Theres a WRONG way to put the objection to the
  half-way house proposal
• This would be to say
• All that a so-called functionally conscious
  system could do is produce the OUTER
  accompaniments to C, rather than the INNER feel

• To talk about phenomenal consciousness as inner
  feel is to be as much in the grip of the thin
  conception as conventional AC research
• The thin conception is
• the idea that phenomenality must be some special
  property, which can be peeled away from all
  external or observable processes
• We have to find an alternative to this idea..

23
The thick conception of consciousness one
version

• Phenomenal consciousness is real and
  non-illusory,
• And it has to be central to any self-respecting
  AC project.
• But it has to be reconceptualized
• by adopting a more grounded and conceptually
  holistic conception of what consciousness is.

• Phenomenal consciousness is important,
• but there is more to consciousness than EITHER
  just phenomenality OR function
• Consciousness is a rich, complex,
  multidimensional cluster of disparate features,
• All (or a significant majority) of these have to
  be seriously taken into account when attempting
  to build consciousness in the way the strong AC
  programme wishes to do.
• The following list is indicative, not definitive

24
Consciousness as a richly interconnected cluster
of features

• Phenomenal feel
• Cognition/intentional content
• Selfhood
• Embodiment/organism
• Neurophysiology
• Architectural/functional aspects
• Affect/teleology
• Behavior/sensorimotor activity
• Environmental coupling
• Intersubjective/developmental aspects
• Ethics/normativity
• Autonomy
• . . . . . ? (an open-ended collection?)

25
Towards a Grand Inventory (i)

• (3) Selfhood
• Self as owner of conscious states
• A perspectival point of view from which states
  are experienced
• Again, there is controversy over whether selfhood
  is more than a formal notion
• - (E.g. T. Metzinger On Being No-one
  conscious minds operate with a transparent
  phenomenal self-model but no real self.)
• (4) Embodiment/organism
• Consciousness as an evolved process how
  ancient?
• Relation between having a metabolism per se and
  inwardness (Hans Jonas)
• The organic-biological nature of all (known)
  consciousness provides a strong challenge to
  computational or robotic forms of AC

• Phenomenal feel (qualia)
• Qualia are the total personal states that we
  are in whenever we have an experience
• E.g. pain isnt just in the foot or in the
  brain (or in the mind)
• The phenomenal feel is not ontologically
  separable from the other features
• its just how it feels to be in that total
  personal state
• (2) Cognition/Intentional content
• Intentionality aboutness e.g. imagining or
  seeing the Eiffel Tower
• Most (all?) phenomenal states are states of
  something i.e. they have intentional content

26
Towards a Grand Inventory (ii)

• Neurophysiological aspects
• Clearly, neurophysiological organisation is
  crucial to known forms of C
• Its become fashionable to search for neural
  correlates of consciousness (F. Crick Youre
  nothing but a pack of neurons)
• Some devotees of consciousness science are
  saying that NCCs should be fully mapped out in 50
  years
• So AC could provide increasingly fine-grained
  computational replications of neural organization
• But will that be sufficient to produce full
  consciousness?
• In any case the goal of mapping NCCs may be
  philosophically problematic

• Architectural / functional aspects
• C clearly plays many roles in our
  information-processing or cognitive functioning,
  e.g.
• attentive direction of effortful tasks
• bringing together disparate processes under
  unified central control
• strategic control of language, perception,
  memory, etc.
• Theres the question of whether these processes
  count on their own as a sort of consciousness?

27
Towards a Grand Inventory (iii)

• (7) Affective/teleological aspects
• Conscious states matter to their bearers, often
  to very large degrees.
• Phenomenology has a positive/negative appraisal
  valency
• Consciousness is bound up with goals only beings
  with intrinsic teleology have phenomenal
  consciousness
• (8) Behavioral/sensorimotor aspects
• Behavior cant be divorced (as 3rd person) from
  consciousness
• Perceptual consciousness as embodied sensorimotor
  activity (ORegan and Noë, Hurley)
• So the behavioral features of consciousness
  arent just (non-conscious) accompaniments, but
  part of the totality of how consciousness
  manifests itself.

• (9) Environmental aspects
• Consciousness cant be properly understood unless
  we include, beside the inner processes
• the subjects lived world,
• and
• the dynamic coupling between subject and world
• (10) Intersubjective (and developmental) aspects
• Our own self-awareness is in large part
  determined by our sense of other selves.
• The way we feel is dependent upon our grasp of
  social dynamics (such as shared attention,
  communicative interaction, theory of mind, etc)
  that we learn from infancy
• Evan Thompson Empathy (experiencing the
  awareness of others) is an precondition of
  consciousness

28
Towards a Grand Inventory (iv)

• Autonomy
• Consciousness is strongly tied up with autonomy
  self-constitution, self-direction, etc.
• Some might say that this is the single most
  important component of consciousness
• And the one that is hardest to implement (in a
  genuine way) within an artificial system
• It obviously connects in many important ways with
  teleology/affect/goals
• Others?
• There are probably several other strands which
  deserve separate mention
• - Self, evolution, ontogeny, autobiography,
  culture, communication, spirituality, etc.

• Normative/moral aspects
• Consciousness is morally critical, as compared to
  other aspects of mind
• Discussions of sentience in non-human animals,
  foetuses, machines, etc. have deep moral
  implications.
• So, too, does AC, in contrast to AI
• to seek to build artificially conscious agents
  raises moral questions that are not raised by
  mere artificially cognitive agents
• So the AC research programme carries a crucial
  ethical responsibility

29
Summing up Three distinctions

• 1. Weak AC versus strong AC
• This discussion has mainly been about the project
  of strong AC
• 2. thin versus thick conceptions of
  (phenomenal) C
• Many proponents of strong AC apply a thin
  conception of C
• as do defenders of traditional (anti-physicalist,
  anti-computationalist) approaches to C
• 3. functional versus phenomenal C
• Its incoherent to claim you have produced the
  first on its own without the second

• Progress in understanding C and AC can best be
  made by exploring a thick conception of C
• Weve done this here by elaborating C as a
  collection of many interconnected constitutive
  strands
• most of these would need to be seriously
  addressed in any properly founded strong AC
  project.

30
Summing up From thin to thick

• Consciousness, defined thinly as just the
  subjective feel (or some other focal key
  property), is a poor foundation for understanding
  consciousness philosophically or scientifically.
• Thin consciousness is also a poor starting
  point for research in artificial consciousness

• By moving to a thick model, AC researchers
  would have a harder, road ahead, but would be
  working within a more realistically grounded
  research framework
• The phenomenon of consciousness will, on the
  thick view, be seen as a diverse network of
  elements, all richly interconnected one with the
  other.

31
Summing up Mixed news

• This is mixed news for a devotee of strong AC
• It shows the task to be much more complex than is
  currently thought.
• Itll require AC researchers to think of
  consciousness in a much more biologically-oriented
  way than they do now
• cf Tom Ziemke,Whats Life got to do with it?
  in recent volume on Artificial Consciousness

• But because of the strong biological orientation,
  
• itll provide a development-path that has
  clearer criteria of success than methods
  currently adopted.

32
(No Transcript)