X-Message-Number: 8663 From: Thomas Donaldson <> Subject: Re: CryoNet #8652 - #8656 Date: Wed, 8 Oct 1997 22:47:06 -0700 (PDT) Hi again! It seems someone else doesn't like the idea of nonTuring Machine computers. As I understand the model set up by the two guys who proposed it, it's not necessary to work in infinite precision. Just work in arbitrarily large precision. So you work with long sequences of integers. In a real device you can do even more. The claim of the paper was that if the weights were incommensurate ie. they were not all rational and therefore could not all be reduced simultaneously to a set of integer weights, then the neural net could not be imitated by a Turing machine. Well, we do have lengths which are incommensurable: the diagonal of a square and the circumference of a circle might be used as such lengths. We can apply rational operations to them, of course, like taking the midpoint. Nor is there any issue of quantum physics which forces space to be digital. Sure, the number of electrons is digital, but space isn't (not that I'm really going to stick on this point --- the reconciliation of general relativity and quantum mechanics still waits to be done, and just what notions we will get out of that I hardly know --- but find it very interesting). For the neural net described in the paper, just one incommensurable weight makes it non Turing. So, if we forget issues of speed, we implement our neural net so that weight is given by (say) the total length of a set of strings kept at each node. And we start with some of these strings given by the length of a diagonal of a square formed by four (other) equal lengths of string. One advantage of neural nets, as parallel machines, is of course that with enough of them you can go fast enough to match any given machine speed. If you don't like the neural net I've suggested because it's too slow, then multiply the number of nodes by 1000 --- or however many. Moreover, since mass == energy/c^2, and energy also need not be digital, then we might even make much better neural nets using a similar idea. Sure, the levels of energy of an electron (say) belonging to an atom are (or might be considered) digital, but then other atoms have other levels, and the energy levels can be affected by nearby atoms too. None of this really affects the abstract theory. I'm just coming up with ideas about how a real device could be a "computer" (whatever a computer is) and still not be a Turing machine. If you want to know my real problem with both Turing machines and the Turing test, it is that I find them both poor models, a poor model for brains and how they work and a poor model for a human being. I've explained in other messages why, and pointed out also that to me this did NOT mean that we could not make a device which would fulfil all our tests for being a human being, at least intellectually. Yes, Searle's ideas did influence me, but then I am not Searle nor do I agree with everything he's said. Welcome to a world in which computers need not be mappable into Turing machines. Best and long long life, Thomas Donaldson Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=8663