Jinhyeok Jeong¹ (), Giwon Bahg¹, Thomas J. Palmeri¹; ¹Vanderbilt University

Ensemble perception is the ability to summarize a collection of objects in a visual array. Most work in ensemble perception has focused on response probabilities. Response times (RTs) have received little attention despite their importance for fully understanding computational mechanisms. We developed a computational modeling framework that provides a process-level account of both response probabilities and RTs in ensemble perception tasks. Our framework allows systematic tests of competing hypotheses regarding how individual objects are perceptually encoded, how ensembles are represented, and how decisions are made. Decisions are modeled as a similarity-based random walk accumulation of evidence, driven by comparisons between a test object and a stored ensemble representation. Given particular assumptions about perceptual encoding and ensemble representations, alternative models can provide alternative mechanistic accounts of how decisions unfold over time, which naturally predicts both response probabilities and RTs. We focus here on three alternative models differing in assumptions about ensemble representation: prototype, summary statistic (mean and variance), and exemplar models. We first evaluated these models by fitting these to an existing dataset in which participants judged the mean orientation of an ensemble array. All three models accounted well for joint distributions of response probabilities and RTs across 7 test level conditions using a single parameter set, demonstrating the viability of the modeling framework. To further differentiate the models, we conducted a new ensemble mean discrimination experiment that factorially manipulated ensemble variance, set size, and test levels, yielding a total of 42 conditions. Summary statistics and exemplar models predicted the major trends in response probabilities and RTs across all conditions, whereas prototype models failed to capture variance effects. These results support a view that ensemble representations are richer than a single abstract mean and illustrate how RTs as well as response probabilities can be accounted for by models of ensemble perception.