Newsgroups: comp.ai,comp.ai.games
Path: cantaloupe.srv.cs.cmu.edu!rochester!cornellcs!travelers.mail.cornell.edu!news.tc.cornell.edu!caen!usenet.cis.ufl.edu!usenet.eel.ufl.edu!news.mathworks.com!gatech!howland.reston.ans.net!ix.netcom.com!netcom.com!nagle
From: nagle@netcom.com (John Nagle)
Subject: Re: WFAQ: TortWorld? (Was: An AI Complementarity Principle)
Message-ID: <nagleDCCoxz.8G8@netcom.com>
Organization: NETCOM On-line Communication Services (408 261-4700 guest)
References: <3v1kb2$ibh@mars.mcs.com> <3v3n7s$gtb@Mercury.mcs.com> <3v5nn0$hji@news.bu.edu> <3v6h6g$jti@Venus.mcs.com>
Date: Thu, 27 Jul 1995 01:17:59 GMT
Lines: 18
Sender: nagle@netcom7.netcom.com
Xref: glinda.oz.cs.cmu.edu comp.ai:31881 comp.ai.games:1931

jorn@MCS.COM (Jorn Barger) writes:
>I haven't really thought about this for neural nets, but I expect that
>a degree of 'reversibility' will be a plausible design strategy for
>NNs in the future.  For example, a neural net chip that generates
>your screen display might be able just as well to do OCR on a scanned
>document-- the same K could and probably should be used for each.

     Neural nets aren't particularly good as a means of building 
reversable models, although you can use one NN to train another
using supervised learning, and build an inverse that way.

     Adaptive model-based feedforward controllers build a model of
the system being controlled in a form that is invertable, which
yields a controller.  Such models are usually constructed with
some form of curve-fitting, using polynomials or splines.  So
there's already a technology that builds and uses reversable models.

				John Nagle
