Richard Bailey¹, Jefferson Ortega¹, Andrey Chetverikov², David Whitney¹; ¹University of California, Berkeley, ²University of Bergen

Ensemble perception allows observers to rapidly extract summary statistics from complex visual scenes. However, it remains unclear whether local correlations within an ensemble bias observers ensemble judgments. Utilizing spatially weighted average models (Pascucci et al., 2021), we investigated whether observers' perceptual judgments in an orientation ensemble judgement task are susceptible to biases towards correlated information in the visual environment across three experiments. In each experiment, participants reported the mean orientation of 36 lines. In Uncorrelated trials, orientations in the ensemble were drawn from the same underlying distribution with a common mean. However, in the Correlated trials, one half of the ensemble display (e.g., left or right) contained orientations sampled from a distribution with a smaller standard deviation than the other side, leading to orientations on one half of the display to have lower variance (i.e., highly correlated orientations). In Experiments 1 and 2, the location of the correlated signal was fixed (always Left or Right for Experiment 1 and 2, respectively). We found that observers had significant spatial biases, consistently over-weighting the side of the display containing the correlated, low-variance information. This occurred regardless of laterality, ruling out inherent visual field anisotropies. In Experiment 3, we investigated whether observers could dynamically shift their spatial weights on a trial-by-trial basis by including both left and right side correlated trials, in addition to uncorrelated trials, interspersed across all trials. Interestingly, we found that participants' spatial weights did not significantly shift towards the correlated side of the display when the correlated signal switched randomly from trial to trial. These results suggest that the visual system is biased towards correlated signals in the environment when making ensemble judgements, but this bias is not dynamic. Instead, observers might update their spatial weights across time through statistical learning, accumulating evidence of environmental regularities over time.