Probabilistic analysis of human supervised learning and classification.

Abstract

Probabilistic classification techniques based on Bayesian decision theory are used to analyze human supervised learning and classification. The procedure rests on the assumption that human classification behaviour is based on internal feature states which can be linked to physical feature vectors (corresponding to the system input). In the present approach, this relationship is modeled in terms of additive stochastic error signals. The corresponding random variables describe the additional degrees of bias and variance introduced by the (perceptual) process of internal feature measurement. Estimates of internal feature states are obtained by least-squares minimization. Structure and dimensionality of the resulting internal representation are displayed by plotting the configuration of internal class means, or virtual prototypes. Their temporal evolution reflects the dynamic properties of the learning process. The use of the procedure is demonstrated by analyzing the results of two experiments. First, it is shown that Minimum Distance Classifiers, such as used by Caelli, Rentschler and Scheidler [(1987) Biological Cybernetics, 57, 233-240], are suboptimal in predicting human performance. Second, it is found that extrafoveal learning is much slower than foveal learning and that extrafoveal pattern representations are severely distorted. The latter distortions reveal the existence of limitations for the generalization of supervised learning over space.