Biased Results with the Leaving-One-Out Method of Pattern Recognition.

Abstract

We show that the leaving-one-out method of pattern recognition must yield biased results when the two sets of training data (representing two classes to be discriminated) are identical. This phenomenon, which we observed during a study of the sensitivity of classification results to errors in the training data, can be eliminated by generating the training sets independently. (Author)

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Dec 11, 1980
Accession Number
ADA096724

Entities

People

  • Kenneth H. Wickwire
  • Lee K. Jones

Organizations

  • Massachusetts Institute of Technology

Tags

DTIC Thesaurus Topics

  • Abstracts
  • Accuracy
  • Air Force
  • Classification
  • Data Sets
  • Errors
  • Estimators
  • False Alarms
  • Kernel Functions
  • Massachusetts
  • Pattern Recognition
  • Reentry Vehicles
  • Sensitivity
  • Test Sets
  • Training
  • United States
  • United States Government

Readers

  • Regression Analysis.

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference