The Role of Categorical and Numerical Reinforcers in Category Learning

Real-world learning signals often come in the form of a continuous range of rewards or punishments, such as receiving more or less money or other reward. However, in laboratory studies, feedback used to examine how humans learn new categories has almost invariably been categorical in nature (i.e. Correct/Incorrect, or A/Not-A). Whether numerical or categorical feedback information leads to better learning is an open question. On one hand, numerical feedback could give more fine-grained information about a category, but may be more uncertain in early learning. On the other, categorical feedback is more dichotomous, potentially leading to larger error signals and more certainty about the outcome. In a series of three studies, the impact of categorical and numerical information was assessed via a multitude of differing category reward structures. To gain a basic understanding of the role that different feedback types have in category learning, Study 1 gave categorical feedback, variable numerical feedback, discrete numerical feedback, and feedback that combined both numerical and categorical information simultaneously to participants who were asked to categorize line stimuli which varied based on two prominent category learning rules. Study 2 expanded on these results and incorporated a basic reward learning manipulation into the task design. In this task, to understand how reward interacts with stimulus similarity, different category clusters were rewarded at different magnitudes with the idea that differences in behavior may arise based on a participant’s sensitivity to either reward magnitude or stimulus similarity. Using a similar paradigm, Study 3 instead altered the rate at which different category clusters were observed. The category and reward learning literatures detail a bias towards stimuli that are more frequent, so this study attempted to determine the potential changes in behavior when stimulus frequency was congruent, or incongruent, with stimulus similarity. The results from each study detail that overall, people seem to learn better from feedback that contains categorical information or rewards that are discrete in magnitude. Further, based on fits of a connectionist model to the behavioral data, people are likely to rely more on stimulus similarity than to any difference in reward magnitude or observational frequency