Path: utzoo!utgpu!watserv1!ssingh From: ssingh@watserv1.waterloo.edu ($anjay "lock-on" $ingh - Indy Studies) Newsgroups: comp.ai Subject: Re: Can Machines Think? Keywords: Neural Networks and Computation Message-ID: <612@watserv1.waterloo.edu> Date: 15 Jan 90 15:52:03 GMT References: <31821@iuvax.cs.indiana.edu> <32029@iuvax.cs.indiana.edu> <1037@ra.stsci.edu> <85217@linus.UUCP> <35@tygra.UUCP> Reply-To: ssingh@watserv1.waterloo.edu ($anjay "lock-on" $ingh - Indy Studies) Organization: U. of Waterloo, Ontario Lines: 48 In article <35@tygra.UUCP> jpp@tygra.UUCP (John Palmer) writes: > >Artificial neural nets will still be unable to solve hard problems >(patttern recognition, REAL language processing, etc) because they >are implemented in silicon (usually as a virtual machine on top of >a standard digital computer) and are therefore inherently inefficient. >In theory (Church-Turing Thesis) it is possible for such problems to >be solved by digital computers, but most of the hard problems are >intractable. We are very quickly reaching the limits of >speed of silicon devices. The inherent inefficiency you attribute to digital computers may be due in part to the Von Neumann bottleneck (See Hillis, W. Daniel, The Connection Machine, MIT Press 1985). The strong version of the Church-Turing Thesis, described in Hofstadter (Metamagical Themas, Bantam Books, 1985) implies that a digital computer can, given enough time, solve ANY and ALL problems. This was the intention of having such a general architecture for computers; ie, if everything is done in the memory, nothing need be physically changed. But when you try to program something as intensive as image or language processing on such a general architecture, things quickly bog down because the serial architecture of the computer, while capable of carrying out the computations necessary, gets stuck in the bottleneck between processor and memory. Image or language processing are problems that lend themselves well to a parallel architecture because they can be broken down and solved over many processors, providing a far greater information throughput than is possible with a purely serial design. Serial machines are good for simulation, because they are so open-ended, but as actual implementations of intelligence, they are somewhat lean. >The only hope of solving these hard problems is by developing devices >which take advantage of the laws of physics and that have a very >strong structure/function relationship. > >My point: We are not going to solve the hard problems of AI by >simply developing programs for our digital computers. We have to >develope hardware that has a strong structure/function relationship. This is why neural nets are the preferred mode of exploration today in large parts of AI research. There does indeed exists a strong structure/ function relation between the NN's parallel design, and the parallel nature of the problems they are being built to solve. -- $anjay "lock-on" $ingh ssingh@watserv1.waterloo.edu "A modern-day warrior, mean mean stride, today's Tom Sawyer, mean mean pride."