Abstract
It has recently been shown that a signal detection model taking into account psychophysical scaling can account for nearly all aspects of visual working memory performance (Schurgin, Wixted & Brady, 2018). A critical prediction of signal detection theory is that when memory is weak, people should frequently ‘false alarm’, sometimes with high confidence. For example, if you are expecting an important call and are therefore carefully monitoring your phone, you might sometimes falsely feel it vibrate, as neural noise in the relevant sensory channel can exceed your threshold for detection. Similarly in memory, looking for weak signals at test (e.g., set size 6) should result in high-confidence false alarms, even to test colors that are categorically distinct from what was encoded. By contrast, if false alarms arise from random guessing, such high-confidence false alarms should not occur, as guessing should be accompanied by low confidence, not high confidence. Across two prominent working memory tasks, we find clear evidence in favor of signal detection. In a change detection task (N=90), we find that participants have a non-trivial number of high-confidence false alarms (and ROCs are curvilinear; see also Wilken & Ma, 2004). In addition, in continuous report tasks (N=30), we also observe high-confidence false alarms, even when accounting for swap errors (e.g., 10% of continuous report errors >90° at the highest confidence level). These results provide strong support for a theory of working memory where noise -- and discerning signal in the presence of noise -- is the limitation on performance, rather than a limit on the number of items that can be represented. Furthermore, we show strong links between change detection and continuous report, arguing that rather than quantifying separate psychological states (“guessing” and “precision”), responses in both tasks are the result of the same underlying signal detection-based process.
Acknowledgement: NSF CAREER BCS-1653457 to TFB