Planning Classification Image Experiments

A. J. Ahumada

NASA Ames Research Center

Abstract

Recent papers on the estimation of classification images can be used to plan experiments. One problem is finding a stimulus domain in which the classification is expected to be primarily linear. A second problem is to decide on the required dimensionality of the domain. A third problem is the type and level of external noise. A fourth problem is estimating the level of internal noise. The final problem is deciding on the size of the confidence intervals or the power for correctly selecting among alternative hypotheses. When all these decisions are made, the number of trials needed can be computed. The paper attempts to help the experimenter make these decisions by exploring alternatives and giving examples.