Long Ni^1,2, Lucas Caceres¹, Michael S. Landy^1,2; ¹Department of Psychology, New York University, ²Center for Neural Science, New York University

Estimates of a stimulus feature (e.g., motion direction) after performing a categorization (e.g,. clockwise vs. counter-clockwise) are repelled from the category boundary (Jazayeri & Movshon, 2007). This phenomenon has been attributed to confirmation bias, where evidence consistent with the chosen category is overweighted (Stocker & Simoncelli, 2007). Is this true for sequential evidence accumulation for categorization? That is, in learning about two categories of stimuli from a sequence of examples, are estimates of the categories themselves biased away from the decision boundary and, if so, is this due to weighting evidence more heavily when it is consistent with the observer’s categorical choice? Eight participants viewed a sequence of 40 ellipses in each block of trials. Ellipse orientation was sampled from one of two partially overlapping Gaussian orientation-category distributions with equal probability. Within a block, the two Gaussian distributions shared the same variance but had different means, equidistant from the optimal decision boundary. Blocks with 8 optimal decision boundaries (0 to 167.5 deg in 22.5 deg steps) were run in random order. Importantly, the decision boundary was never explicitly provided to participants and thus was learned while performing the task. Participants reported the stimulus category and received auditory feedback about correctness. At the end of the block, participants estimated the mean orientation of each category. Participants’ estimates were substantially biased away from the decision boundary. Regression analysis showed that stimulus orientations that were correctly categorized contributed strongest to the mean estimates, consistent with confirmation bias. However, after controlling for the magnitude of orientations, we found that the repulsive bias was not due to confirmation bias but, rather, to overweighting stimulus orientations farther away from the optimal decision boundary. Our findings suggest that repulsive bias can occur in the absence of confirmation bias and thus highlight the need to consider alternative mechanisms.

Acknowledgements: NIH EY08266