Browse Prior Art Database

Adaptive Pattern Recognition

IP.com Disclosure Number: IPCOM000090475D
Original Publication Date: 1969-Apr-01
Included in the Prior Art Database: 2005-Mar-05
Document File: 3 page(s) / 53K

Publishing Venue

IBM

Related People

Chow, CK: AUTHOR

Abstract

On the basis of a set of sample patterns, whose features optionally can be discrete or continuous, the adaptive pattern recognition system adjusts its structural and recognition weights to better reflect the structure of the input pattern classes.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 55% of the total text.

Page 1 of 3

Adaptive Pattern Recognition

On the basis of a set of sample patterns, whose features optionally can be discrete or continuous, the adaptive pattern recognition system adjusts its structural and recognition weights to better reflect the structure of the input pattern classes.

The system evaluates the mean and variance of each feature and co- variances of pairs of these features. It computes the correlation coefficients between feature pairs. It uses the relative magnitudes of these coefficients as branch weights to construct a statistical tree of dependence. The tree and its associated parameters then are stored for future recognition.

In the drawing, the expression x = (x(1), x(2)... x(n)) denotes the features of the pattern, where n is the number of features. Feature x(i) can be continuous or discrete. The number of states x(i) need not be finite.

The recognition function, for the pattern class k, is of the form.

(Image Omitted)

where b, w(1), w(2), u and v are recognition weights and the set {j(i,k)} represents the dependence tree.

Once these parameters are determined, the recognition function can be readily implemented, as explicitly indicated in the above Equation, by a set of difference-squaring devices and a weighting and summing network. The number of squaring devices is at most equal to the number of features. This represents no increase over the implementation for independent normal distributions.

The crucial problem is to adapt the recognition weights and the tree-structure to the input data. For simplicity of description, only one pattern class is considered, and the index k is disreg...