A computational modeling framework for ensemble perception

Ensemble perception refers to the ability to summarize a collection of objects in a visual array. Implicit ensemble perception tasks ask participants to judge if a test object belongs to a studied visual array, with a bias towards the central tendency of an array used as a measure of ensemble representations. Explicit ensemble perception tasks ask participants to compare a test object with the mean of a studied array, with performance used as a measure of the ability to extract ensemble statistics of an array. We developed a computational modeling framework to understand ensemble perception that allowed us to systematically test alternative competing hypotheses about model mechanisms regarding how individual objects in a visual array are perceptually encoded, how ensembles are represented, and how decisions in both implicit and explicit ensemble tasks are made. Perceptual encoding was implemented with several forms of capacity limitation, including different levels of subsampling, noisy encoding, and memory strength scaling with set size. Ensemble representations were implemented with different levels of abstraction, either as an abstract mean representation or a set of exemplar representations. Implicit vs. explicit ensemble decisions were implemented as different forms of similarity-based random walk accumulation of evidence. We tested whether specific models instantiated in this framework could account for a wide range of ensemble perception phenomena from both implicit and explicit tasks. For implicit tasks, models predicted a central tendency bias and poor recognition of individual objects regardless of assumptions made about perceptual encoding and ensemble representations. For explicit tasks, models predicted set-size effects only when perceptual encoding assumed noise, and predicted variance effects only when ensemble representations contained more than just an abstract representation. Our modeling framework provides a theoretical space for testing hypotheses regarding computational mechanisms involved in ensemble perception.