Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP Path: utzoo!watmath!clyde!rutgers!princeton!mind!harnad From: harnad@mind.UUCP Newsgroups: comp.ai,comp.cog-eng Subject: Re: Minsky on Mind(s) Message-ID: <479@mind.UUCP> Date: Tue, 3-Feb-87 11:10:05 EST Article-I.D.: mind.479 Posted: Tue Feb 3 11:10:05 1987 Date-Received: Sat, 7-Feb-87 11:52:39 EST References: <463@mind.UUCP> <464@mind.UUCP> <2099@dciem.UUCP> <471@mind.UUCP> <2117@dciem.UUCP> Organization: Cognitive Science, Princeton University Lines: 127 Keywords: processes, consciousness, epiphenomenalism Summary: When and why man/machine similarities are relevant: The difference between the informal and the scientific problem of inferring mind Xref: watmath comp.ai:206 comp.cog-eng:54 mmt@dciem.UUCP (Martin Taylor) writes: > we DO see some things as more alike than other things, because > we see some similarities (and some differences) as more important > than others. The scientific version of the other-minds problem -- the one we deal with in the lab and at the theoretical bench, as opposed to the informal version of the other-minds problem we practice with one another every day -- requires us to investigate what causal devices have minds, and, in particular, what functional properties of those causal devices are responsible for their having minds. In other words (unless you know the answer to the theoretical problems of cognitive science and neurosience a priori) it is an EMPIRICAL question what the relevant underlying functional and structural similarities are. The only defensible prior criterion of similarity we have cannot be functional or structural, since we don't know anything about that yet; it can only be the frail, fallible, underdetermined one we use already in everyday life, namely, behavioral similarity. Every other similarity is, in this state of ignorance, arbitrary, a mere similarity of superficial appearance. (And that INCLUDES the similarity of the nervous system, because we do not yet have the vaguest idea what the relevant properties there are either.) Will this state of affairs ever change? (Will we ever find similarities other than behavioral ones on the basis of which we can infer consciousness?) I argue that it will not change. For any other correlate of consciousness must be VALIDATED against the behavioral criterion. Hence the relevant functional similarities we eventually discover will always have to be grounded in the behavioral ones. Their predictive power will always be derivative. And finally, since the behavioral-indistinguishability criterion is itself abundantly fallible -- incommensurably moreso than ordinary scientific inferences and their inductive risks -- our whole objective structure will be hanging on a skyhook, so to speak, always turing indistinguishable from state of affairs in which everything behaves exactly the same way, but the similarities are all deceiving, and consciousness is not present at all. The devices merely behave exactly as if it were. Throughout the response, by the way, Taylor freely interchanges the formal scientific problem of modeling mind -- inferring its substrates, and hence trying to judge what functional conditions are validly inferred to be conscious (what the relevant similarities are) -- with the informal problem of judging who else in our everyday world is conscious. Similarities of superficial appearance may be good enough when you're just trying to get by in the world, and you don't have the burden of inferring causal substrate, but it won't do any good with the hard cases you have to judge in the lab. And in the end, even real-world judgments are grounded in behavioral similarity (indistinguishability) rather than something else. > it is simpler to presume that Ken and Steve experience > consciousness than that they work according to one set of > natural laws, and I, alone of all the world, conform to another. Here's an example of conflating the informal and the empirical problems. Informally, we just want to make sure we're interacting with thinking/feeling people, not insentient robots. In the lab, we have to find out what the "natural laws" are that generate the former and not the latter. (Your criterion for ascribing consciousness to Ken and me, by the way, was a turing criterion...) > All the TTT does, unless I have it very wrong, is provide a large set of > similarities which, taken together, force the conclusion that the tested > entity is LIKE ME The Total Turing Test simply requires that the performance capacity of a candidate that I infer to have a mind be indistinguishable from the performance capacity of a real person. That's behavioral similarity only. When a device passes that test, we are entitled to infer that its functional substrate is also relevantly similar to our own. But that inference is secondary and derivative, depending for its validation entirely on the behavioral similarities. > If a simpler description of the world can be found, then I no > longer should ascribe consciousness to others, whether human or not. I can't imagine a description sufficiently simple to make solipsism convincing. Hence even the informal other-minds problem is not settled by "Occam's Razor." Parsimony is a constraint on empirical inference, not on our everyday, intuitive and practical judgements, which are often not only uneconomical, but irrational, and irresistible. > What I do argue is that I have better grounds for not treating > these [animals and machines] as conscious than I do for more > human-like entities. That may be good enough for everyday practical and perhaps ethical judgments. (I happen to think that it's extremely wrong to treat animals inhumanely.) I agree that our intuitions about the minds of animals are marginally weaker than about the minds of other people, and that these intuitions get rapidly weaker still as we go down the phylogenetic scale. I also haven't much more doubt that present-day artificial devices lack minds than that stones lack minds. But none of this helps in the lab, or in the principled attempt to say what functions DO give rise to minds, and how. > Harnad says that we are not looking for a mathematical proof, which is > true. But most of his postings demand that we show the NEED for assuming > consciousness in an entity, which is empirically the same thing as > proving them to be conscious. No. I argue for methodological epiphenomenalism for three reasons only: (1) Wrestling with an insoluble problem is futile. (2) Gussying up trivial performance models with conscious interpretations gives the appearance of having accomplished more than one has; it is self-deceptive and distracts from the real goal, which is a performance goal. (3) Focusing on trying to capture subjective phenomenology rather than objective performance leads to subjectively gratifying analogy, metaphor and hermeneutics instead of to objectively stronger performance models. Hence when I challenge a triumphant mentalist interpretation of a process, function or performance and ask why it wouldn't function exactly the same way without the consciousness, I am simply trying to show up theoretical vacuity for what it is. I promise to stop asking that question when someone designs a device that passes the TTT, because then there's nothing objective left to do, and an orgy of interpretation can no longer do any harm. -- Stevan Harnad (609) - 921 7771 {allegra, bellcore, seismo, rutgers, packard} !princeton!mind!harnad harnad%mind@princeton.csnet