Item

Abstract

The aim of the study was to find out how humans learn to use informative features to categorise novel objects. Subjects were presented with a sequence of shapes, which they had to learn to classify into two categories. Immediate feedback about the true category was given after each trial. Stimuli consisted of large irregularly shaped contours containing several protrusions ('features'), whose curvatures varied stochastically from trial to trial. The exemplars for each class were drawn from partially overlapping Gaussian distributions in a multi-dimensional feature space. Thus, a single feature alone was often not sufficient to discriminate between classes and exemplars could be ambiguous. The features were independent and varied in diagnosticity (d') and perceptual discriminability. Importantly, the stimulus design made it possible to use eye-movement recordings to measure which features subjects looked at to perform the task. The results show that humans can learn to discriminate stochastic categories despite the inherent ambiguity of the task, and that with increasing expertise the fixations become more focused possibly reflecting the subject's belief about relevant features. The human data (performance, reaction times, and eye movements) are compared with an ideal observer and a variety of rational (Bayesian) learner models.