I knew that! Response-based Outcome Predictions and Confidence Regulate Feedback Processing and Learning

Abstract

Influential theories emphasize the importance of predictions in learning: We learn from response outcomes and feedback to the extent that they are surprising, and thus convey new information.Here we investigated how individuals learn to predict response outcomes based on the subjective confidence and objective accuracy with which these predictions are made. We hypothesized that both of these aspects modulate how feedback is processed and that they are reflected in eventrelated potentials (ERPs) as measured using EEG. Participants performed a time estimation task with graded, performance-contingent feedback. With this design we could distinguish reward prediction errors (RPE), indexing outcome valence with regard to the goal, and output prediction errors (OPE), indexing the absolute mismatch between predicted motor outcome and actual performance. As we expected, predictions made with higher confidence were more accurate (smaller OPE), and more so as learning progressed. Further, individuals with a better correspondence between confidence judgments and prediction accuracy learned more quickly. Outcome valence, as indexed by RPE was reflected in the feedback-related negativity (FRN). In contrast, P3a amplitude increased with OPE and confidence, that is with the degree of surprise about the outcome. Finally, performance-relevant information converged in the P3b component with confidence modulating RPE effects in early trials while learning took place. Taken together, the results underline the significance of different aspects of predictions and suggest a role of confidence in learning. CONFIDENCE IN PREDICTIONS AND FEEDBACK regulation of the decision process both within and between individuals (Bahrami et al., 2010; Shea et al., 2014). Both of these forms of performance evaluationerror detection and confidence (Yeung & Summerfield, 2012)require sufficient knowledge about the task at hand.However, little is currently known about how people learn to make these evaluations as they master new skills, and this is particularly true in the context of tasks involving continuous response parameters-as in motor control, e.g., when throwing at a target-where errors are inevitable and graded, as compared with the simpler case of a binary, categorical choice. CONFIDENCE IN PREDICTIONS AND FEEDBACK 4In the present study we used this framework of confidence and error detection as a form of outcome prediction to investigate how individuals learn to accurately evaluate responses along two key dimensions: 1) predicting action outcomesi.e., how people learn to accurately predict the direction and magnitude of a potential error, and 2) accurately indicating confidencei.e., how people learn to calibrate their confidence judgments such that higher confidence correlates with more accurate predictions. We further studied how these internal evaluations affect the encoding of feedback information as it is reflected in event-related potentials (ERPs) of the EEG.

When effort matters: Expectations of reward and efficacy guide cognitive control allocation

“…We found a different pattern of results when examining the feedback-related negativity (FRN), which typically indexes the difference in feedback-locked activity for trials that resulted in negative compared to positive feedback 46 . Consistent with previous findings [47][48][49][50][51][52] , we found a reliable effect of receipt vs omission of reward on FRN amplitude (b = 0.80, p <.001), and this effect was enhanced for high reward trials (b = 0.81, p = .007; Table S9). However, in addition to this, and contrary to the hypothesis we preregistered based on previous findings 53, 54 , we found that effects of reward receipt vs omission on FRN amplitude were reduced for trials with high efficacy compared to those with low efficacy (b = -0.83, p = .007; Fig.…”

“…Epochs containing artifacts, with amplitudes exceeding ±150 µV or gradients larger than 50 µV, were excluded from further analysis. We focused our analyses on these event-related potentials, quantified agnostic of condition with ROIs and time-windows determined a priori based on the literature : cue-locked P3b (250 to 550 ms, averaged across Pz, P3 and P4 27,83 ), cue-locked late contingent negative variations (CNV; to 1500 ms post-cue, i.e., -500 to 0 ms pre-target, averaged across Fz, FCz, and Cz 30 ), response-locked correct-and error-related negativities (CRN/ERN; 0 to 100 ms 43,84 ), and feedback-locked feedback-related negativity (FRN; quantified peak-to-peak at FCz as the difference between the negative peak between 250 to 350 ms and the positive peak in the preceding 100 ms from the detected peak 85 ). All EEG data preprocessing were performed using custom MATLAB scripts using EEGLAB functions (cf.…”

Motivation and cognitive control in depression

“…The priors were very wide Gaussian distributions for both the intercept (M = 5; SD = 5) and the slopes (M = 0; SD = 5; cf. Fischer & Ullsperger, 2013;Frömer, Nassar, et al, 2018).…”

“…All three models included intercept as a constant and a varying effect, and constant main effects in the varying effect structure. We selected wide Gaussian priors based on previous studies (Fischer & Ullsperger, 2013;Frömer, Nassar, et al, 2018) for both the intercept (M = 0; SD = 5) and the slopes (M = 5; SD = 5).…”

The Global Workspace Needs Metacognition

“…Subjective confidence is used to weight prior knowledge and incoming evidence by their respective reliability [110]. Individuals with a better correspondence between confidence judgments and prediction accuracy learn more quickly [119]. Individuals with better metacognitive accuracy for a perceptual decision are also better able to learn novel arbitrary cue-stimulus associations for the stimuli the perceptual decision was based on [120].…”