X-Message-Number: 11887 From: Thomas Donaldson <> Subject: For Bob Ettinger and others: computers, symbols, and reality Date: Sat, 5 Jun 1999 00:24:40 +1000 (EST) Hi to everyone! About computers and symbolic computation: one subtle point is easy to forget. No, the lowest level operations of our brain are not symbolic in any meaningful sense... although we certainly operate with symbols when we think (most people think in a mixture of images and words). The key here is the definition of "symbol". A symbol is an ARBITRARY thing or event which is associated with some other real thing or event. When our brain works at the lowest level (ie. before we get into language) it is not working with symbols because the brain events are not arbitrary. Our DNA, for instance, is not symbolic. It consists of a string of biochemicals with a relationship to other systems defined by chemistry, and that relationship does not hold for arbitrary sets of biochemicals. As for our thinking, I distinguish the thinking of which we are conscious from that which actually goes on --- and sometimes finishes even before we become conscious of it. After all, consciousness is sequential and our brain is very highly parallel --- more parallel than any computer yet built. A lot of our thinking about visual scenes and even relations between things, animals, and people does not happen on a symbolic level. We simply have systems which respond to events because of their structure, not their components ... and that structure most certainly cannot be arbitrary. And if we think about how language REALLY works, ultimately there is an association of a symbol with a thing, event, etc which is NOT learned simply through a set of definitions of symbols. This is why neural nets provide a much better foundation for "intelligent" (and perhaps even conscious) robots than do ordinary computers. Ordinary computers merely operate with symbols; we human beings, using those computers, decide on what their results mean. An ordinary computer cannot do this, no matter how powerful. Neural nets, however, come much closer to linking the operations of the neural net computer to those of the world in a way which is NOT symbolic. And yes, because ordinary computers can be programmed to tell us things in ordinary language it is easy to think that their language works the same way as ours --- which is quite false. The one issue which Searle does not discuss, and which remains an open question, is the one I've discussed before: Can a symbolic system be sufficiently involved and complex that it can have ONLY ONE possible interpretation in the world? Basically that would make the symbolic system NOT arbitrary. I personally strongly doubt that. But to prove that this is possible or impossible requires much more than the kind of handwaving that normally happens --- and has not yet been done either way. Basically, I am saying that Searle had the beginning of a key point. More understanding of how brains actually work emphasizes this more and more. For those who want devices which don't work like ordinary computers, but which ARE intelligent, this simply means that they should not use ordinary computers to make their devices. (Even with lots of add-ons they will still fall down, basically because NO system of symbols can ever come close to the real world). And if we want these devices to approach the abilities of brains we may even have to use neural nets unlike those now used: NNs that grow and change constantly, even at the level of adding new neurons (a side comment: our brains probably make new neurons, just like those of many other animals, in two regions AT A MINIMUM: the dentate gyrus of our hippocampus and the neurons of our olfactory system. The new neurons come not by division of old neurons, but from stem cells which grow into neurons). And for computer people there is an extra kink with neural nets: it's far from obvious that they can be imitated by Turing machines. The reason I suggest this is simple: classical Turing machines respond to symbols on a tape by moving the tape and changing the symbols. Those neural nets able to learn on their own (not all neural nets can do this) DEPEND on sensors telling them about the world, which is not a tape to be moved or written upon. I expect that some readers of this will answer by providing plans by which a Turing machine might get inputs (the inputs to a Turing machine ie. what is on the tape at the start, are not normally specified) and copy a neural net in its behavior. I raise a suggestion here, not provide any kind of proof; but I will point out that those inputs to a computer trying to simulate a neural net will almost inevitably be symbols rather than real events (the Turing machine has a tape, not eyes or ears). And best and long long life to everyone, Thomas Donaldson Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=11887