Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!rutgers!att!cbnewsl!apr
From: apr@cbnewsl.ATT.COM (anthony.p.russo)
Newsgroups: comp.ai.neural-nets
Subject: Re: : Step Function
Summary: X-OR is NOT learnable
Message-ID: <1667@cbnewsl.ATT.COM>
Date: 29 Aug 89 11:40:29 GMT
References: <1060@rex.cs.tulane.edu> <6980@sdcsvax.UCSD.Edu> <6981@sdcsvax.UCSD.Edu>
Organization: AT&T Bell Laboratories
Lines: 22

In article <6981@sdcsvax.UCSD.Edu>, demers@beowulf.ucsd.edu (David E Demers) writes:
> 
> >Some things have been proven not to be learnable, i.e. EX-OR.
> 
> Surely you jest.  XOR is not learnable by a neural net with
> no hidden layers, but is certainly a learnable function.
> 

X-OR is not learnable. If you are given the first three entries in the truth
table, you could not possibly generalize to the last entry with
any confidence. That is, of the two possibilities for the last entry,
neither is preferable.

X-OR is *memorizable*, not learnable.

In either case, learnability is not a requirement for neural net applicability.
Nets memorize well.

	~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
	~  	 Tony Russo		" Surrender to the void."	~
	~   apr@cbnewsl.ATT.COM						~
	~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~