Abstract
Images can be represented in visual working memory (VWM) not only on a pixel-by-pixel basis, but also in a more abstract way, in terms of their statistical structure. We previously showed (ARVO, 2002) that for 1st-order statistics (luminance) and local 4th-order statistics (higher-order structure), this representation is graded, not categorical. The goal of this study is to examine the timecourse of pixel-by-pixel and statistical representation of images in VWM.
Stimuli exhibited three kinds of statistical changes: changes in a 1st-order statistic (luminance), changes in a local 4th-order statistic (isodipole structure), and changes in a non-local statistic (presence or absence of mirror symmetry). In a 4-AFC task, subjects (N=5) viewed four 8x8 arrays of black and white checks and were asked to determine which one (the target) changed from S1 (600ms) to S2 (200ms). A blank ISI was set to either 200 or 1000ms in alternating blocks of trials. In half of the trials, 16 checks within the target changed in luminance between S1 and S2, without producing a change in image statistics. In the other half of the trials, the change in the target from S1 to S2 induced a change in its statistical structure.
Across all subjects and stimulus conditions, fraction correct (FC) and reaction time (RT) improved with a change in statistics (FC increased by .20, RT decreased by 144ms). For local statistics, a larger improvement in performance was observed for the 1000ms ISI condition (1st-order: .39 FC increase, 321ms RT decrease; 4th-order: .22 FC increase, 173ms RT decrease) than in the 200ms ISI condition (1st-order: .29 FC increase, 156ms RT decrease; 4th-order: .19 FC increase, 102ms RT decrease). For mirror symmetry there was only a minimal effect: .10 FC increase, 81ms RT decrease at 1000ms; .04 FC increase, 56ms RT decrease at 200ms ISI.
We conclude that the representation of image statistics in visual working memory plays a larger role with a longer retention interval.