A Signal-Detection Model of Visual Short-Term Memory Performance for Random Block Patterns

Abstract

On the basis of signal-detection theory, we have formulated a model that accurately explains performance on a visual short-term memory task involving random block patterns. The model assumes that the internal response of an observer for detecting a change in any given element of the block pattern is noisy and has a Gaussian-shaped distribution. On this basis we can calculate the likelihood that an observer correctly or falsely identifies a change in the pattern after a certain time interval (ISI). Using this likelihood, we can then predict the likelihood that an observer correctly identifies a whole pattern as having changed or not as a function of the number of elements that changed in the pattern. We have previously shown (Cornelissen and Greenlee, 1993 Perception22 Supplement, 46) that memory performance declines when changes occur in pattern elements located on the perimeter of the pattern. Therefore the model also incorporates a circular symmetric ‘memory field’ that shows a Gaussian-shaped decline of memory performance from the point of fixation. The model has three parameters: d' (detectability of a change), lambda (criterion level), and the standard deviation of the Gaussian of the memory field. In the experiments we performed, block patterns made up of 50 light and 50 dark randomly arranged elements (0.5 deg checks) were briefly (200 ms) shown. In a forced-choice task, subjects judged whether two sequentially presented (with ISIs of 1, 3, or 10 s) block patterns were the same or different. Task difficulty was varied by varying the number of elements in the patterns that changed on ‘different’ trials. The model is able to accurately predict memory performance at the three different ISIs for various levels of pattern differences (changes in 1, 2, 4, 8, 16, 20, and 50 out of 100 elements).