The following is a summary Mahesh prepared of an invited talk he
saw at IICAI (http://www.iiconference.org/) It raises a number of
interesting issues that are worth thinking about. Comments or
questions are of course welcome.
======================================================================
Is Pattern Recognition a Statistical Problem?
Dr. Yegnanarayana (IIT Madras)
The talk strongly argued about the difference between pattern recognition
and statistical problems.
On a spectrum that ranges from highly pattern recognition type problems
(towards left end) to very statistical type problems (towards right end),
Dr. Yegnanarayana placed signature matching, image recognition etc. on
extreme left, handwriting recognition and speech recognition as somewhere in
between and problems like population analysis, stock market prediction etc.
on the extreme right as highly statistical problems (where there is
probabilistic distribution involved, but not a pattern).
An appealing example he used was that of showing the audience an outline of
a bird and then gradually adding the details. With the first outline itself
one could recognize a bird in the picture and the question he asked was
what statistics was involved in our brain recognizing that image as that of
a bird? No statistical analysis of the pixel intensities was involved surely
(as most of such information was missing in the picture which was simply an
outline). This is what he said was pattern recognition.
Pattern is a multi-dimensional multi-modal feature. It implicitly exists in
the relations among the various variables.
Pattern recognition involves dealing with a *small amount* of
*high*-dimensional or multi-variate data. The key lies in the relation among
these multiple variables.
On the other hand, a statistical problem deals with *large amount* of
*low*-dimensional data. Here the key lies in the distribution of the
features.
What we call as the curse of dimensionality in pattern recognition problems
might indeed contain valuable relationships, which define the pattern.
Therefore dimensionality reduction can throw away the pattern, which we are
trying to find.
The way a human does pattern absorption is due to *permissible variations*
i.e. the generalization property of human mind (e.g. we don¹t fail to
recognize an owl as a bird if we initially know only a crow as a bird). So
pattern absorption is not due to some statistical probability distribution
that our mind has formed, but due to these permissible variations that we
are flexible to.
Our speech recognition and understanding is based on selective attention and
does not involve isolated stages of processing as in first pre-processing,
then analysis etc. But it is more of an integrated process.
Overall it was a very thought-provoking talk and made us think as Dr.
Yegnanarayana said are we looking under the lamp post for the lost wallet
just because there is light available under it, whereas the wallet is
actually somewhere in the dark? i.e., are we using statistical methods for
pattern recognition just because they are the means available to us, and the
real solution is something else?
Yahoo! Groups Links
<*> To visit your group on the web, go to:
http://groups.yahoo.com/group/nlpatumd/
<*> To unsubscribe from this group, send an email to:
[EMAIL PROTECTED]
<*> Your use of Yahoo! Groups is subject to:
http://docs.yahoo.com/info/terms/
