Teruaki Kido¹, Yuka Sasaki¹, Nitzan Censor^2,3, Takeo Watanabe¹; ¹Brown University, Providence, USA, ²Sagol School of Neuroscience, Tel Aviv University, ³School of Psychological Sciences, Tel Aviv University

Perceptual learning has traditionally been induced and studied through repetitive practice. The texture-discrimination task (TDT; Karni & Sagi, 1991) is a standard training paradigm for perceptual learning research, yet several procedural and analytical aspects may limit the accuracy with which learning magnitude is estimated. Here, we examined methodological factors that may have influenced earlier findings and propose modifications to improve the reliability of TDT-based learning measures. Specifically, we evaluated three potential issues: (1) imperfect gaze control, (2) task complexity during the response phase, and (3) analyses potentially prone to bias, which could lead to undesirable gaze strategies, non-perceptual improvements, or estimation bias in task performance. First, we investigated how the lapse rate might affect performance measures. Specifically, we simulated data computationally while varying the underlying lapse rate, then compared the performance estimates from different lapse-rate modeling approaches. We found that conventional approaches could significantly overestimate performance at relatively high lapse rates, which may occur in the TDT due to its dual-task nature. Next, we conducted a single-session behavioral experiment with the TDT to test procedural and analytical modifications. We monitored gaze, preventing its deviations in real time, and simplified the response procedure. To reduce bias, we added catch trials with long stimulus-to-mask onset asynchronies and used a bias-robust model (Prins, 2012). When data and analysis aligned with those of previous studies, the estimated task performance was consistent with that in the preceding studies, confirming task compatibility despite the procedural modifications. However, with catch trials and the bias-robust method, performance appeared worse than in earlier findings, suggesting overestimation in those studies. Together, these findings underscore the importance of improved data modeling and procedural control. Implementing the proposed modifications may enable more accurate estimation of perceptual learning magnitude in TDT-based experiments.

Acknowledgements: NIH R01EY027841, NIH R01EY019466, NIH R01EY031705, and BCS-2241417