By Topic

Classification with finite memory

Sign In

Cookies must be enabled to login.After enabling cookies , please use refresh or reload or ctrl+f5 on the browser for the login options.

Formats Non-Member Member
$31 $13
Learn how you can qualify for the best price for this item!
Become an IEEE Member or Subscribe to
IEEE Xplore for exclusive pricing!
close button

puzzle piece

IEEE membership options for an individual and IEEE Xplore subscriptions for an organization offer the most affordable access to essential journal articles, conference papers, standards, eBooks, and eLearning courses.

Learn more about:

IEEE membership

IEEE Xplore subscriptions

2 Author(s)
Wyner, A.D. ; AT&T Bell Labs., Murray Hill, NJ, USA ; Ziv, J.

Consider the following situation. A device called a classifier observes a probability law P on l-vectors from an alphabet of size A. Its task is to observe a second probability law Q and decide whether P≡Q or P and Q are sufficiently different according to some appropriate criterion. If the classifier has available an unlimited memory (so that it can remember P(z) exactly for all z), this is a simple matter. In fact for most differentness criteria, a finite memory of 2(log A)l+o(l) bits will suffice (for large l), i.e., store a finite approximation of P(z) for all Alz's. In a sense made precise in this paper, it is shown that a memory of only about 2Rl bits is required, where the quantity R<log A, and is closely related to the entropy of P. Further, it is shown that if instead of being given P(z), for all z, the classifier is given a training sequence drawn with a probability law P that can be stored using about 2Rl bits, then correct classification is also possible

Published in:

Information Theory, IEEE Transactions on  (Volume:42 ,  Issue: 2 )