Using Ideal Observers in Higher-order Human Category Learning

Abstract

Ideal observer models have proven useful in investigating assumptions about human information processing in a variety of perceptual tasks. However, these models have not been applied in the area of higher-order category learning. We describe a simple Bayesian ideal observer and apply it to empirical data on category learning. We describe an experiment in which we found that acquisition of family resemblance categories was drastically impaired if the categories were defined by relations between features rather than by the features themselves. An ideal observer was used to test whether this effect could be accounted for by inherent information differences between the conditions. A comparison of participants’ performance to the model found a significant difference in efficiency of learning even after accounting for information differences between conditions. This analysis illustrates how ideal observer methods can provide useful tools for analyzing higher-order category learning.