Abstract
Visual Working Memory (VWM) performance may emerge from factors such as the likelihood that a given item has been processed (Pm) and the resolution of stored representations (SD). Recent approaches to modeling VWM have assumed that, when an item is not processed (i.e., Pm=0), observers’ guesses will be uniform – but the actual pattern of guesses was not measured. And, has assumed that internal representations are subject to Gaussian noise (SD) that is constant across various feature values – but this constancy was not empirically tested. These assumptions make generalization difficult when guesses are not uniform (e.g., biased guessing) and when internal noise varies as a function of the parameter value (e.g., scalar variability in numerosity; Whalen, 1999; oblique effect in orientation; Appelle, 1972). Here we demonstrate a general-purpose approach to modeling performance in VWM tasks that includes: 1) empirically measuring guesses on trials with 0ms-stimulus-duration, 2) generating appropriate models of internal representation for particular feature dimensions, and 3) iteratively generating mixture models to estimate a likelihood that each observation derives from an internal representation. In three experiments, we use this method to estimate Pm and SD for color, orientation, and numerosity of arrays of 1, 3, or 6 sets at stimulus duration of 0, 33, 67, 100, 133, or 198ms. Results show that Pm increases from 0 to 100ms then attains asymptote while SD remains stable across all SOA's, suggesting that information accumulation in perception completes within 100ms and that stored information in VWM is all-or-none. SD remained constant across set sizes, favoring a fixed-resolution model of VWM. This approach embraces and extends previous modeling efforts, allowing us to empirically estimate guessing, capture a wider range of appropriate models for internal representations, and analyze data at the individual trial level.
Meeting abstract presented at VSS 2012