‘What a Mistake!’: Prediction error modulates explicit and visuomotor predictions in virtual reality

Contemporary Bayesian theories of the mind highlight the role of prediction errors as a driving force in learning processes and the updating of our models of the world and of ourselves. This process is postulated to manifest both at explicit levels (e.g., subjective responses) and at implicit levels (e.g., eye movements, reaction time) of behavior. We developed a novel associative learning task in immersive virtual reality, in which a cue probabilistically predicted the subsequent location of a target to which participants had to reach. Specifically, the current study (N = 30) examined: (1) whether explicit predictions (i.e., responses) of target location corresponded to implicit expectations and gaze patterns. (2) how prediction errors on the previous trial updated participants’ expectations and behavior in the subsequent trial. First, we found strong and consistent evidence that explicit predictions were closely linked to eye movements, such that dwell time was significantly greater in the expected location of the upcoming target (p < .001, Cohen’s d = 1.57). Second, following erroneous trials there was a modulation in both explicit and implicit levels of behavior. Thus, following an error, explicit responses were more likely to be switched (p < .001, Cohen’s d = 2.05) reaction time was significantly longer (p < .001, Cohen’s d = 0.86), and the degree by which the current trial’s dwell time relied on the current cue significantly decreased (p < .001, Cohen’s d = 1.12). These changes presumably reflect the updating of the subject’s model of the environment following an error. To further examine this updating, we modeled the trial-by-trial prediction error derived from the Rescorla-Wegner model of reinforcement learning. Preliminary results found that the degree to which the current trial’s gaze direction relied on the cue was significantly correlated to the previous trial’s prediction error. This effect was robust both at an individual subject level (17/30 participants had significant correlations) and at a group-level (p < .001, Cohen’s d = 1.48). Using this paradigm, we aim to examine the effect of prediction error on additional implicit measures (e.g., pupillometry, hand kinematics), and develop unique markers of prediction error modulation. These markers may be altered in psychopathological conditions such as psychosis in which prediction errors are postulated to be aberrant.