Path: utzoo!attcan!utgpu!watmath!att!dptg!rutgers!network!sdcsvax!beowulf!pluto
From: pluto@beowulf.ucsd.edu (Mark E. P. Plutowski)
Newsgroups: comp.ai.neural-nets
Subject: Re: : Step Function
Message-ID: <6996@sdcsvax.UCSD.Edu>
Date: 31 Aug 89 04:22:41 GMT
References: <1060@rex.cs.tulane.edu> <6980@sdcsvax.UCSD.Edu> <17522@bellcore.bellcore.com> <1683@cbnewsl.ATT.COM>
Sender: nobody@sdcsvax.UCSD.Edu
Reply-To: pluto@beowulf.UCSD.EDU (Mark E. P. Plutowski)
Organization: EE/CS Dept. U.C. San Diego
Lines: 17

In article <1683@cbnewsl.ATT.COM> apr@cbnewsl.ATT.COM (anthony.p.russo) writes:
>Perhaps our definitions of "learnable" are different. Mine is that,
>with a fraction of the possible samples, one can generalize to 100%
>accuracy.

I would suggest you introduce one of two things into your 
definition of "learnability."   Either 1) make it dependent on
a particular architecture,  or, 2) make it dependent upon a
complexity measure, such as time needed to learn, number of units,
or number of weights.  

It makes sense to discuss learnability when we talk about a set
of concepts (e.g., functions) given a hypothesis space with which
to learn the concepts (e.g., threshold-logic).  Then, it is an
issue whether the concept in question can be represented within
the hypothesis space -- and then, we can discuss whether it can
be learned given a feasible amount of resources.