SANG HUN LEE¹ (), Dong-gyu Yoo; ¹Seoul National University

The prevalence of cardinal orientations in our visual surroundings appears to be reflected in our perceptual system, possibly influencing our estimation performance in two error metrics. When assessed for bias, orientation estimates are typically displaced away from their nearest cardinal orientations, often referred to as the “repulsive bias.” When assessed for variability, orientation estimates are typically more variable across trials for oblique orientations than cardinal orientations, often referred to as the “oblique effect.” While these two phenomena have been widely reported, they have been studied in quite an isolated manner. Consequently, little effort has been made to relate the orientation-specific bias to the corresponding variability. Given the intimate interdependence between these two error metrics in the estimation theory, such as the Cramer-Rao bound, clarifying and accounting for the relationship between the orientation-specific bias and variability will refine the current understanding of how the visual system operates to estimate the state of its surroundings. In this perspective, we gathered the orientation estimation data from four independent studies with different viewing conditions and examined whether any significant relationship exists between the orientation-specific bias and variability. Here, we report one such relationship: as the repulsive effect grows in the bias, the oblique effect retreats while the anti-oblique effect (i.e., cardinal effect) begins to dominate the variability. When we quantified the strength of the repulsive bias and the dominance of the oblique effect in the variability, their relationship was well captured by a linear trend across the 10 different datasets (r = -0.948), and across individual observers within the datasets (r = [-0.618, -0.703]). By leveraging the “efficient Bayesian observer model” proposed by Wei and Stocker (2015), we attempted to offer a principled account for the observed bias-variability covariance, which points to the variation in “loss function” as a computational-level source of the covariance.

Acknowledgements: This work was supported by the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT (No. NRF-2015M3C7A1031969, NRF-2017M3C7A1047860, NRF-2021R1F1A1052020), and SNU R&DB Foundation (339-20220013).