Estimates of the Expected Quality of Signs,

Abstract

In constructing pattern recognition algorithms a certain number of characteristic features, which are selected during the learning process, are used. In selecting such features it is intended that the probability of the recognition error be minimized. Therefore, the problem of evaluating the quality of selected features arises, which is analyzed in this article. Two methods which make it possible to select the most informative features are presented. The first method is based on the selection of the class of features for a set of examples used in the learning process. It is shown that the probability of a recognition error using that feature (the quality of a feature) is a monotone function of given parameters. The second method based on the fact that during the learning process the approximate given distribution of the probabilistic characteristics of features can be constructed. The given distribution is the a priori distribution of the probabilities that the analyzed feature has a given measure and that the pattern having that feature belongs to a given class. This distribution is further used in determining the quality of selected features.

Document Details

Document Type
Technical Report
Publication Date
Nov 05, 1970
Accession Number
AD0717887

Entities

People

  • M. M. Bongard
  • M. N. Vaintsvaig

Organizations

  • National Air and Space Intelligence Center

Tags

Communities of Interest

  • Materials and Manufacturing Processes

DTIC Thesaurus Topics

  • Algorithms
  • Identification
  • Learning
  • Mathematics
  • Monotone Functions
  • Pattern Recognition
  • Probability
  • Recognition

Readers

  • Artificial Intelligence
  • Regression Analysis.

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference