Inferring the order of stable and context dependent perceptual biases in human vision

Perception is transformed by both stable and time varying (contextual) expectations. This influence is observable in the domain of orientation processing as cardinal bias (repulsion from more common vertical and horizontal orientations) and serial dependance (attraction towards recent stimuli) [1,2]. These biases operate on vastly different time scales and have typically been studied in isolation. Investigating the properties of these two biases in tandem may reveal important insights into general principles of perception (efficient coding, Bayesian inference) that are obscured when studied in isolation. Here, we modeled the responses of human observers in a 2AFC delayed orientation discrimination task. Responses displayed substantial cardinal (6.5±1.2° peak repulsion from cardinal axes, mean±SD) and serial biases (3.6±1.9° peak attraction towards previous stimulus) relative to their overall precision (σ=9.9±0.36°). By leveraging incorrect responses, we found that the center of attraction towards the previous trial was heavily shifted towards the mis-remembered item (shift 16.79±0.79°). To further infer the order that expectations are applied in this hidden process, we fit 6 alternative models and evaluated the resulting cross-validated likelihoods. Responses were best explained when the inducing stimulus included all known biases (including cardinal, its own history bias, and residual misperception not accounted for) and this dynamic history information was integrated at the end of the processing stream (p<.001 compared to all competing models). This suggests that while long term priors are used to make the earliest stages of encoding more efficient, short term expectations are applied at a later stage to support immediate goals. More generally, these results provide a framework for understanding how long term priors (likely arising during development) interact with dynamic contextual factors to jointly drive complex behaviors. [1] Girchick et al.,2011. NatNeuro; [2] Fischer&Whitney,2014. NatNeuro