Path: utzoo!attcan!uunet!cs.utexas.edu!usc!ucsd!cogsci!meadors
From: meadors@cogsci.ucsd.EDU (Tony Meadors)
Newsgroups: comp.cog-eng
Subject: Re: More flexible handwriting recognition?
Message-ID: <765@cogsci.ucsd.EDU>
Date: 16 Aug 89 07:06:17 GMT
References: <1440003@hp-ptp.HP.COM>
Reply-To: meadors@cogsci.UUCP (Tony Meadors)
Organization: U.C. San Diego, Institute for Cognitive Science
Lines: 41

In article <1440003@hp-ptp.HP.COM> garye@hp-ptp.HP.COM (Gary_Ericson) writes:
>In the handwriting recognition systems that I have seen, the user is required 
>to enter the handwritten characters into a set of boxes.  This isolates the 
>characters and reduces deviations in size and orientation.  I'm wondering if 
>any work has been done to allow more flexibility.  Specifically,
>
>	1. Allow characters to be entered anywhere in the workspace (not just
>	   in pre-defined locations).
>
>	2. Allow characters to be created in any size and, possibly, have 
>            ...
>	3. Allow characters to be entered in any orientation and, possibly,
>            ...
>Anyone know of research being done along these lines 
>            ...
    The ability to recognize the letter "A" in all "valid" instantiations
    has been called the most central issue and challenging problem for
    artificial intelligence (a paraphrase of Douglas Hofstadter). The 
    implication intended is that this problem of "interpretation" is a
    microcosm of all the larger issues AI must tackle.

Commentary:
    Maybe Hofstadter is on target about AI. But with regard to an "adaptive,
    behaving, intelligence" we might do better to consider the central
    problem to be more, "how to place the peg in the proper hole
    problem."

In any event, the universal recognition of letters has been recognized
as important, and the subject of much "research", for some time. But the
problem remains very much a challenge.

Of course a system might be engineered which satisfactorily performs the
task of letter classification (including the three requirements you 
mention), yet have little relevance to AI *or* cognitive psychology
(ie., the method might be applicable only to letter classification, and
be nothing like human classification). Yet AI reseachers have on the
whole been quite willing to try just about any sort of algorithm...if 
it only would do the job: it's a pesky, non-trivial problem.

tonyM