Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!rutgers!tut.cis.ohio-state.edu!gem.mps.ohio-state.edu!csd4.csd.uwm.edu!bionet!agate!shelby!lindy!news From: GA.CJJ@forsythe.stanford.edu (Clifford Johnson) Newsgroups: comp.ai Subject: Re: Are neural nets stumped by change? Message-ID: <4356@lindy.Stanford.EDU> Date: 15 Aug 89 17:28:31 GMT Sender: news@lindy.Stanford.EDU (News Service) Distribution: usa Lines: 49 In article, jk3k+@andrew.cmu.edu (Joe Keane) writes: >In article (Clifford Johnson) writes: >> In adapting to change, they [NNs] only do so according to >>statistical/numerical rules that are bounded by their (implicit >>or explicit) preprogrammed characterizations and >>parameterizations of their inputs. > >Some neural networks have carefully hand-crafted topologies. But if you use a >standard topology and training algorithm in a new domain, where is the >``preprogramming''? That's why I was careful to state "implicit or explicit" re the "preprogramming." Whatever the topology, a definite set of distribution functions is implied. True, convergence of recognition outputs to fit a very wide range of inputs may be engineered, but convergence takes time. It proceeds in steps determined by the topology, and assumes a constant sampling space. The lack of constancy, i.e. change, is what stumps it. >Similarly, with a standard topology, you aren't giving it >any ``parameterization''; it learns them all by itself. Yes and no. The parameter-space is basically bounded by the topology. You can't, for example, have more degrees of freedom learned by the system than exist in its topology. And again, a change in the external or real parameters is only relearned over time, which is my main point. >>Thus, a change in the basic >>*type* of pattern is beyond their cognition. > >This doesn't follow. It may seem intuitive to you, but i think it's false. >Fill in some more steps and i'll tell you where i think the problem is. If a neural-net optical character reader is suddenly confronted with chinese characters, it isn't going to learn to read them, if it's only classification choices are arabic. Continued training might result in a systematic many-to-one translation of chinese characters into their "closest" arabic equivalents, closest being dependent on the aforesaid net's topological design. Yes, a better net might be built to include the capability to develop further classifications (again this takes time), but it wouldn't have a clue as to what the new patterns "mean" in terms of decision-making that was originally defined only in arabic terms.