Visual category learning with dimensionally-separable stimuli : a comparison of performance between pigeons and humans.

Understanding how organisms learn perceptual categories on the basis of experience has been an important goal for researchers in a number of subdisciplines of psychology, including behavior analysis, experimental psychology, and comparative cognition. The primary aim of this thesis is to investigate how nonhumans (pigeons) and humans learn to make visual category judgments when stimuli vary quantitatively along two dimensions, particularly when accurate responding requires integration of information from both dimensions. The thesis consists of four chapters and a technical appendix. Chapter 1 is a literature review which provides a broad overview of studies on categorization by nonhumans and humans, as well as specific background for the current research. Chapters 2 and 3 constitute the empirical portion of this thesis. Four experiments are described, using a category task based on the ‘randomization’ procedure developed originally by Ashby and Gott (1988) with human participants and employed in subsequent research by Ashby, Maddox and their colleagues (see Ashby & Maddox, 2005; Maddox & Ashby, 2004, for review). Stimuli were Gabor patches that varied in frequency and orientation. Our primary goals were to determine whether pigeons could respond accurately in an information integration task with dimensionally-separable stimuli, and to compare performances of pigeons and humans. Chapter 2 reports two experiments with pigeons. Experiment 1 compared performance in two conditions which varied in terms of whether accurate performance required control by both dimensions (“information integration; II) or by a single dimension (“rule based”; RB). Results showed that pigeons learned both category tasks,with an average percentage of correct responses of 85.5% and 82% in the II and RB conditions, respectively. Although perfect performance was possible, responding for all pigeons fell short of optimality. Model comparison analyses showed that the General Linear Classifier (GLC; Ashby, 1992), which has been proposed to account for category learning in similar tasks with humans, provided a better account of responding in the II conditions, but a unidimensional model that assumed control only by frequency provided a better account of results from the RB condition. Thus results show that pigeons can respond accurately in an information integration task based on dimensionally-separable stimuli. However, analysis of residuals showed that systematic deviations of GLC predictions from the obtained data were present in both II and RB conditions. Specifically, accuracy for one category (A) was an inverted-U shaped function of orientation, whereas accuracy for the other category (B) did not vary systematically with orientation. Results from the RB condition showed evidence of an interaction between frequency and orientation, such that accuracy was higher for orientation values that were relatively low (i.e. close to horizontal) than high (i.e., close to vertical). Experiment 2 compared responding in two RB conditions which differed in terms of whether frequency or orientation was the relevant dimension. Pigeons again responded accurately in the task. Results from the frequency-relevant condition replicated the interaction obtained in Experiment 1, whereas results from the orientation-relevant condition gave no evidence of an interaction. Chapter 3 reports two experiments which compare performances of pigeons (Experiment 1) and humans (Experiment 2) in category tasks using identical stimuli. In each experiment there were two conditions, both based on the information-integration task in which the range of orientation values was wide or narrow. There were two primary goals. First, we wanted to test whether the inverted-U shaped pattern for Category A accuracy as a function of orientation would be replicated with different pigeons and stimulus values. Second, we wanted to compare responding of pigeons and humans. A secondary aim was to test whether restriction of range would affect control by orientation. Results from the condition with a wide orientation range were similar to those from Chapter 2, and showed that the inverted-U shaped pattern was replicated for both pigeons and humans. When the range of orientation values was narrow, responding for both pigeons and humans was exclusively controlled by orientation. Overall, results for pigeons and humans were similar and suggest that a common process may underlie information-integration category learning in both species. Chapter 4 provides a summary of the empirical results from Chapters 2 and 3, and shows that the inverted-U shaped pattern of accuracy for Category A as a function of orientation is unanticipated by current models for category learning, such as the GLC, prototype theory, and exemplar theory. A new ‘fuzzy prototype’ model is described which provides a good account of the results and predicts the inverted-U shaped pattern. According to the new model, subjects associate a linear segment in the stimulus space (‘fuzzy prototype’) with one of the category responses. When a stimulus is presented on a trial, subjects are assumed to use an ‘A/Not-A’ decision rule, with the probability of a Category A response determined as a function of the minimum distance of the stimulus from the fuzzy prototype. Possible directions for future research are considered. The thesis concludes with a technical appendix which describes the experimental chambers, interface hardware, and computer software developed to conduct the research,and a detailed user’s manual for the software. The system allows the same control procedure for both human and pigeon experiments, and should be useful for future research on categorizatio