Path: utzoo!utgpu!utstat!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!rutgers!rochester!pt.cs.cmu.edu!cadre!geb From: geb@cadre.dsl.PITTSBURGH.EDU (Gordon E. Banks) Newsgroups: comp.ai Subject: Re: letter to THE NEW YORK REVIEW concerning AI Keywords: Intention Message-ID: <2411@cadre.dsl.PITTSBURGH.EDU> Date: 12 Mar 89 22:59:07 GMT References: <7471@venera.isi.edu> <2447@crete.cs.glasgow.ac.uk> <509@mmlai.UUCP> <2481@crete.cs.glasgow.ac.uk> <2369@cadre.dsl.PITTSBURGH.EDU> <2564@crete.cs.glasgow.ac.uk> Reply-To: geb@cadre.dsl.pittsburgh.edu (Gordon E. Banks) Organization: Decision Systems Lab., Univ. of Pittsburgh, PA. Lines: 40 In article <2564@crete.cs.glasgow.ac.uk> gilbert@cs.glasgow.ac.uk (Gilbert Cockton) writes: > >comment. As far as I know, there is only one way to "train" a neural >network, whereas the growth of understanding in the presence of an >expert passes through several forms of training, which may eventually >reduce to the shake of a head and a gesture. > >Is there a sense in which neural network training requires > > a) an artificial, well-designed task > b) continued practice over this task. > >If so, then ho, ho, ho, because life just ain't so simple. The good >guys in this world are the ones who don't need Skinnerian programmed >instruction. > >The topic is life, not the behavioural modification of the mentally ill >or deficient. There are several ways to train neural networks, and we are continually discovering new ones. None of the ways are likely to be the same as the way real neural networks learn (yet). Another way to make a neural network is to hard-wire it so that it does it tasks well from the beginning. The "good guys" like you and me who aren't mentally deficient had a lot of our connections already made when we came out of our mammas. Figuring out what those connections are is the province of the neuroanatomists as well as the congnitive scientists. Neural network training often produces some surprises even to the trainers. For example, if you look at Hinton's and Rumelhart's program where they trained a network to distinguish the letter C from the letter T (letters could be rotated in 90 degree increments), some of the internal neurons developed as center surround detectors without being told to do so. These detectors are found in retina and brain of all species at least down to the toad. We aren't ready to build a human brain yet (still working on the toad), but I think this is right track (rather than using production rules). We can't be expected to recapitulate a billion years of evolution in 20.