Computationally Efficient Estimators for the Bayes Risk.
Abstract
A computationally efficient estimator for the Bayes risk is one which achieves a desired accuracy with a minimum of computation. In many problems, for example speech recognition, point evaluations of the class conditional densities are computationally costly. Density evaluations are the single most important factor contributing to the computational effort in Bayes risk estimation, thus the amount of computation required by a Bayes risk estimator is defined as the average number of conditional density evaluations it performs. The accuracy of risk estimator is defined by its variance. In practice, the true optimal estimator cannot be determined since this would in effect require knowledge of the true risk R. Thus a technique whereby the first n of the total N samples are used to approximate the optimal estimator is proposed. The n samples should contain enough information on the closeness of the classes to determine an almost optimal estimator. The last N-n samples are used in the approximate optimal estimator to obtain an accurate estimate of the risk with a minimum of computation.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 1978
- Accession Number
- ADA055997
Entities
People
- Lynn D. Wilcox
- Rui J. P. De Figueiredo
Organizations
- Rice University