Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!europa.eng.gtefsd.com!howland.reston.ans.net!EU.net!uunet!olivea!news.hal.COM!decwrl!netcomsv!netcom.com!park
From: park@netcom.com (Bill Park)
Subject: Re: Successful application of neural nets  
Message-ID: <parkCy75pD.3r5@netcom.com>
Followup-To: comp.ai.neural-nets
Summary: Is CATPAC a handwriting/printing/script recognition system?
Keywords: CATPAC Newton neural network handwriting recognition
	learning artificial intelligence
Organization: Netcom Online Communications Services (408-241-9760 login: guest)
References: <Cy2quM.BJC@acsu.buffalo.edu>
Date: Mon, 24 Oct 1994 21:44:00 GMT
Lines: 53

In article <Cy2quM.BJC@acsu.buffalo.edu> comjoew@ubvms.cc.buffalo.edu
(Joseph D Woelfel) writes:

> 	Arbitron uses a neural net called CATPAC to read thousands
> 	of radio listerner's diaries on a regular basis.Johnson
> 	controlsw uses it to read qualitative responses of customers
> 	about their performance; all the top level execs and managers
> 	use it, including engineering. Wherever large amounts of text
> 	are in use, CATPAC can usually be used effectively.
> 
> 	Joseph Woelfel
> 	University at Buffalo

Thank you for your reply to my query about successful applications of
neural networks.  CATPAC sounds like a good example.

But I am not sure I understand what CATPAC does, exactly.  In
particular, what kind of text does it read?  What do Arbitron diary
entries or Johnson Controls responses look like? Does CATPAC read
absolutely anything a person may write in such documents, or does it
only recognize a small set of easily-distinguishable words chosen for
the specific application (e.g., radio station call letters, or
qualitative assessments such as "good," "medium," "bad?").  Mark-sense
forms are another rather low-tech way to collect such information:
Draw a line with a soft pencil in an indicated place on a paper form
corresponding to your response; detect the mark optically or
electrically; position of the mark on the form encodes the response
information. Widely used in academic testing.  Does CATPAC have some
crucial advantage over that much simpler technology?

Does CATPAC read ordinary handwritten words, phrases, and sentences?
If so, that would be remarkable, to say the least: Consider the
disappointing performance of the Apple Newton, for example. Although
it is generally regarded as one of the more successful attempts to
read unconstrained handwriting and printing mixed together, it has
received no end of criticism for its error rate.  It was even
lampooned in the Doonesbury comic strip!  Yet the Newton's task is
much simpler than CATPAC's in several important ways: The Newton only
has to read a single person's handwriting.  It is allowed to learn
how that person writes (you correct it when it misinterprets what you
wrote; takes a week or two to learn well enough to be useful).
Finally, the Newton has access to the writing dynamics in the form of
the time history of the stylus motion.  CATPAC would have to work only
with the final, static image, without any dynamic information.

If CATPAC can actually read static handwriting images with high
accuracy, a large lexicon, and no training on samples of handwriting
from the writer, then whoever sells CATPAC ought to call the Apple
Newton Group immediately!

Bill Park
=========
-- 
Grandpaw Bill's High Technology Consulting & Live Bait, Inc.
