2.5 |. Modeling PDR response using a learning model

We used a learning model to examine the trial-by-trial changes of prediction and prediction error in the task. We chose the RW learning model because it measures how prediction error affects the strength of predictions. Therefore, the model will be sensitive to the strength of the predictions our participants have for the visual stimuli, as they view present trials (designed to strengthen predictions) and omission trials (designed to generate prediction error).

We assumed individual PDR at each trial reflects the magnitude of prediction error or surprise in that trial (Kloosterman et al., 2015; Lavin et al., 2014; O’Reilly et al., 2013; Preuschoff et al., 2011; Sirois & Jackson, 2012). For each participant, we fitted a generative implementation of a learning model to their PDR response. The model asserts that prediction error is calculated for each trial, t:

where: O (t) is the appearance (or omission) of visual stimulus at trial t and P (t) is the predicted probability of that appearance. In addition, predictions are updated for each trial based on the prediction error and learning rate, where α is the learning rate:

We fitted the individual models by minimizing the Mean Square Error (MSE) between the PDR and δ² (level of surprise or magnitude of prediction error regardless of the direction of the error) throughout the experiment for each participant (see Supporting Information 2.5). Furthermore, fitting the model revealed each participant’s learning rate, α, and initial prediction, P (0).