Abstract
After viewing a series of sequentially presented visual stimuli, subjects can readily generate mean representations of various visual features. Although these average representations seem to be formed effortlessly and obligatorily, it is unclear how such averages are actually computed. According to conventional prototype models, the computation entails an equally weighted average taken over all the stimuli. To test this hypothesis, we had subjects estimate the running averages of some feature in a series of sequentially presented stimuli. Part way through the series, we perturbed the distribution from which stimuli were drawn, which allowed us to test alternative models of the computations behind subjects' estimates. In both explanatory and predictive tests, a model in which the most recent items had disproportionate high weight outperformed a model in which all items carried equal weight. Such recency-weighted behavior was shown consistently in multiple experiments in which subjects estimated running averages of length of vertical lines. However, the degree to which recent items were prioritized varied with the type of stimulus, such that when estimating the running averages of a series of numerals, subjects showed less recency prioritization. We conclude that previous evaluations of prototype models have made unrealistic assumptions about the nature of a prototype, and that a reassessment of prototype models of visual memory and perceptual categorization may be in order.
