nach oben

Minds and Machines

Erschienen in:

01.11.2012

How the Problem of Consciousness Could Emerge in Robots

verfasst von: Bernard Molyneux

Erschienen in: Minds and Machines | Ausgabe 4/2012

Einloggen

Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.

search-config

KI-gestützte Suche

Aus

Abstract

I show how a robot with what looks like a hard problem of consciousness might emerge from the earnest attempt to make a robot that is smart and self-reflective. This problem arises independently of any assumption to the effect that the robot is conscious, but deserves to be thought of as related to the human problem in virtue of the fact that (1) the problem is one the robot encounters when it tries to naturalistically reduce its own subjective states (2) it seems incredibly difficult from the robot’s own naturalist perspective and, most importantly, (3) it invites the robot to engage in the exact same metaphysical responses as humans offer to the problem of consciousness. Despite the fact that it invites the robot to consider extravagant metaphysical solutions, the problem I explore is purely algorithmic. The robot cannot complete its naturalist physicalist reduction as a matter of algorithmic fact, whether or not the naturalist physicalist reduction would be correct as a matter of metaphysical fact. It is hoped that by reproducing the familiar seeming problem in an artificial context, a greater understanding of the human problem of consciousness can be achieved.

Nächster Artikel Thinking Inside the Box: Controlling and Using an Oracle AI

Sie haben noch keine Lizenz? Dann Informieren Sie sich jetzt über unsere Produkte:

Springer Professional "Wirtschaft+Technik"

Online-Abonnement

Mit Springer Professional "Wirtschaft+Technik" erhalten Sie Zugriff auf:

über 102.000 Bücher
über 537 Zeitschriften

aus folgenden Fachgebieten:

Automobil + Motoren
Bauwesen + Immobilien
Business IT + Informatik
Elektrotechnik + Elektronik
Energie + Nachhaltigkeit
Finance + Banking
Management + Führung
Marketing + Vertrieb
Maschinenbau + Werkstoffe
Versicherung + Risiko

Jetzt Wissensvorsprung sichern!

Jetzt informieren

Springer Professional "Technik"

Online-Abonnement

Mit Springer Professional "Technik" erhalten Sie Zugriff auf:

über 67.000 Bücher
über 390 Zeitschriften

aus folgenden Fachgebieten:

Automobil + Motoren
Bauwesen + Immobilien
Business IT + Informatik
Elektrotechnik + Elektronik
Energie + Nachhaltigkeit
Maschinenbau + Werkstoffe

Jetzt Wissensvorsprung sichern!

Jetzt informieren

Springer Professional "Wirtschaft"

Online-Abonnement

Mit Springer Professional "Wirtschaft" erhalten Sie Zugriff auf:

über 67.000 Bücher
über 340 Zeitschriften

aus folgenden Fachgebieten:

Bauwesen + Immobilien
Business IT + Informatik
Finance + Banking
Management + Führung
Marketing + Vertrieb
Versicherung + Risiko

Jetzt Wissensvorsprung sichern!

Jetzt informieren

We give M ‘his' own arbitrary gendered pronoun for readability.

We can grant, if intentional content requires consciousness, that these tokens have content only from our human perspective. I.e. that they are merely truth evaluable, or otherwise semantically evaluable, by us.

Not the converse (and more controversial) identity of indiscernibles, for which the term 'Leibniz's Law' is also sometimes used.

Obviously, the conclusion that an epistemically diligent M-machine cannot complete the task entails the conclusion that if an M-machine completes the task, it was not epistemically diligent. An M-machine could therefore have a belief in the identity, even a correct belief, but, assuming that such things require epistemic diligence, not justified belief or knowledge.

Since nth order properties always apply to n-1th order phenomena.

Strictly speaking, nothing in the following requires us to agree with M's arguments. To derive M's hard problem, we do not need something real that M cannot reduce: We only need something that M believes in.* We could therefore rewrite the following so as to put the existence of a single subjective correspondent into escrow.

* Of course, if the subjective correspondent lacks existence then that explains why M cannot find any objective phenomenon with which to correctly identify it. However, that would be the uninteresting explanation. Where Sherlock Holmes seems the kind of thing that would be reducible if he existed, conscious experience does not. So if subjective correspondents are to be more like conscious experiences than Sherlock Holmes, we need to show them to be the kinds of things that would seem irreducible on the assumption they exist.

There is an intuitive notion—a possible analog of M's commitment—that we humans cannot be wrong about our own conscious states. Alternatively, M's commitment might correspond to 'taking consciousness seriously'—i.e. to not 'redefin[ing] the phenomenon in need of explanation as something it is not' (Chalmers 1996, x).

The fallibilist about conscious experience might, consistent with her view that we can mistake pain for pressure in the dentist's chair, nonetheless concede that some experiences are experiences we can't be wrong about. E.g., I cannot be overruled by a doctor when I think I am in agonizing pain. Similarly, a robot without M's infallibillist streak might nonetheless express certainty concerning the nature of certain subjective states, With respect to such states, a hard problem can be derived.

The wrinkle in allowing M such infallible authority is that, in some cases, his behaviors may diverge. E.g., asked to press a button only if he seems to see a red flash, M might, through some processing malfunction, press the button while verbally denying that he saw the flash. Similar cases occur in the clinical literature on brain damaged humans. The right response, I think, is not to treat them as counterexamples to infallibility, but to treat them as cases where the agent herself is fragmented.

Given that it does not matter whether M is right or wrong about his own infallibility, one might wonder why I spend so much time emphasizing the plausibility of M’s stance. It is because I wish to forestall complaints to the effect that M’s problem arises from his adopting an unreasonable position, or from his being unlike humans. If it is granted, though, that M's position is not weird, and that it resembles our own to the appropriate extent, then, even if it is incorrect, we have enough to press forward.

An identification can be said to be unfinished, in the sense we mean, if it is nonconstructive or if the identification of its properties, or its property's properties, etc., is nonconstructive. A restriction that prohibits nonconstructive identifications thereby prohibits unfinished ones.

Nor, for the same reason, could M have already identified greenishness with some property that already applies to C.

For Xⁱ to seem some way or other is to seem discernible from the other ways it might have seemed, a truth that holds even if it is a property. But Xⁱ can only seem discernible from other ways it could have seemed if it itself seems to have properties. X seems discernible from Y, after all, only if the properties of X themselves seem discernible from those of Y.

A slightly different argument to the same conclusion: Either Xⁱ is a property or not. Suppose not. In that case, since it seems some way or other, it must seem to have some color, be some size and shape, and so on. Hence Xⁱ seems to have properties. If, on the other hand, Xⁱ is some seeming property of Xⁱ⁻¹ then Xⁱ itself must seem to have the property of modifying X ⁱ⁻¹; of being apparent; of being a property; of marking X ⁱ⁻¹ as distinct from things that lack X ⁱ; etc. It will, moreover, have idiosyncratic properties too. E.g., if Xⁱ is the property (say) of being green, it will seem to have the property of marking its bearer as colored; of making its bearer visually distinguishable from red things; of camouflaging its bearer against green backdrops; and so on. So again, if Xⁱ seems some way, then it seems to have properties.

Moreover, M infers that it does. For M is positioned to know that Xⁱ seems to M to have Xⁱ⁺¹—or else it does not really seem that way to M after all—and M can infer from this, together with his rule PI, that SXⁱ must have SXⁱ⁺¹. Since we are assuming that M infers the properties of potential identificanda whenever he is in a position to, it follows that he does.

See e.g., Dennett (1988), Rey (2007).

See Jackson (1982) and Chalmers (1996) for recent property dualist positions.

To get something closer still to Cartesian dualism, M could also pursue the same strategy in response to the corollary, holding that he is not the hardware that he sees in the mirror, nor any other objective thing he finds in the physical world.

It might be thought that even when relaxing the third restriction M might prove, using just the first two restrictions, that no specific identification is possible in principle, hence hankering after a future finished theory would make no sense. This tempting thought is only half right. M could make specific but unfinished identifications by identifying A with B while only committing to each of A's properties being identical to some property or other of B. In this way a non-specific identification at order O + 1 permits an unfinished but specific identification at O.

This tendency to halt and leave the matter unfinished, typically at a very low order, may plausibly explain why the human problem of consciousness does not obviously present itself as a regress.

Compare, obviously, with the question: 'How could pain possibly be nothing more than c-fibers firing?' Identifications which proceed while accompanied by such mystification are dubbed 'gappy' identifications by Levine (2001), who regards them as the typical reaction to phenomenal-physical identifications. Such ‘how possibly’ reactions, I suggest, arise from the fact that the identification hasn't been properly squared with Leibniz's Law.

Paradoxes illustrate that (apparently) deductive arguments only rationally compel belief when the conclusion is appropriately acceptable. Zeno's paradox, for example, was an apparently sound deductive argument to the effect that bodies could not move. Those who historically struggled to deny either the premises or deductive validity of the argument were nonetheless rational to resist its conclusion.

Subtraction is another option. To identify heat with molecular motion, for example, humans may either add the 'feeliness' of the heat as a new property of the molecular motion or just as plausibly remove the 'feeliness' from the heat by re-categorizing it as a property of the human's own sensory system, not of the heat itself. Since M claims no infallibility when it comes to objective phenomena like heat, he can, consistent with his first restriction, revise away those of its properties that make it difficult to reduce.

Contrast this with how M (or a human) is able to identify (say) the morning star with the evening star. Assuming they only ever differed with respect to one property, after adding the property appears in the morning to the evening star and the property appears in the evening to the morning star the identification can go ahead. Why is there no regress? Because all the specific properties of properties of… the morning star and properties of properties of… the evening star end up being the same, to infinity, when the differing properties are added. To see why, let the evening star's 'property tree' be the structure of its properties, its properties' properties etc., branching up to infinity. Now take an arbitrary property P* that appears in this property tree somewhere in the sub-tree above the property P that was added to the morning star. Since P was added, the whole of the sub-tree to which it leads is added too, and so P* appears in the morning star's property tree after all. (see Fig 3.).

Indeed, without a premise like (3), M has not even got the weaker argument to the effect that his subjective correspondents are identical to some functional state or other of the motherboard, or CPU, etc.

I thank Matthias Scheutz for pushing the possibility of an ingenious proof, as well as the possibility of a looping property space. Note, though, that no proof of a single identity is sufficient to escape the problem—see objection 1. What's needed is a proof of infinitely many specific identities, and the mind boggles at what that might look like. We should also be pessimistic about the possibility of a repeating property space, since the property is a property of a property of a property of a red thing cannot logically appear at any level lower than the 3rd; and there are many properties of this sort at every order (including some more interesting ones like is the horribleness of the hunger-feeliness of the biological state of a badger).

A programmer who knew the specific identities to infinity might enter a rule that generates them for M. Even given such a wonderful rule, though, the identification still appears intractable. For M must still identify the properties PX and PY of X and Y (using the rule) before identifying X and Y, and the properties PPX and PPY of PX and PY (using the rule again) before identifying those, and so on. There is no starting point to such a process even if each identification takes only one step.

The programmer could just give M the same representation for both identificanda thus saving it from even having to identify them. However, in that case the robot would not be able to contemplate them being distinct, since it needs separate representations to do that. We would have yet another machine that dodges its hard problem by being insufficiently smart.

Indeed, it is perhaps because we each take on of these strategies for avoiding the regress and try to make it work (e.g., by taking the dualist option and then struggling with mental causation) that explains why we get the problems associated with the ways out of the regress, not the problems associated with the regress itself; hence why it is that the problem of consciousness does not seem to present itself as a regress.

Chalmers, P. (1996). The conscious mind. Oxford: Oxford University Press.MATH

Dennett, D. (1988). Quining qualia. In A. Marcel & E. Bisiach (Eds.), Consciousness in modern science. Oxford: Oxford University Press.

Jackson, F. (1982). Epiphenomenal qualia. Philosophical Quarterly, 32, 127–136.CrossRef

Levine, J. (2001). Purple haze: The puzzle of consciousness. Oxford: Oxford University Press.CrossRef

Minsky, M. (1965). Matter, mind and models. Proceedings of International Federation of Information Processing Congress, 1, 45–49.

Nagel, T. (1974). What is it like to be a bat? Philosophical Review, 83(4), 435–450.CrossRef

Rey, G. (2007). Phenomenal content and the richness and determinacy of colour experience. Journal of Consciousness Studies, 14(9–10), 112–131.

Titel: How the Problem of Consciousness Could Emerge in Robots
verfasst von: Bernard Molyneux
Publikationsdatum: 01.11.2012
Verlag: Springer Netherlands
Erschienen in: Minds and Machines / Ausgabe 4/2012
Print ISSN: 0924-6495
Elektronische ISSN: 1572-8641
DOI: https://doi.org/10.1007/s11023-012-9285-z

Springer Professional

Abstract

Bitte loggen Sie sich ein, um Zugang zu Ihrer Lizenz zu erhalten.

Sie haben noch keine Lizenz? Dann Informieren Sie sich jetzt über unsere Produkte:

Springer Professional "Wirtschaft+Technik"

Springer Professional "Technik"

Springer Professional "Wirtschaft"

Weitere Artikel der Ausgabe 4/2012

Hitoshi Iba, Yoshihiko Hasegawa, and Topon Kumar Paul: Applied Genetic Programming and Machine Learning

Thinking Inside the Box: Controlling and Using an Oracle AI

Probabilistic Belief Contraction

The Explanatory Role of Computation in Cognitive Science

Premium Partner