Improving task performance through practice can occur at various levels, including action selection, decision processes, and stimulus perception. Explicit feedback significantly enhances learning across these levels. However, improvements also occur without explicit feedback, suggesting the role of metacognitive processes like confidence. Confidence, an estimate of decision accuracy, can act as a proxy for explicit feedback ('implicit feedback'). Current understanding of confidence's role in learning lags behind that of explicit feedback. Reinforcement learning models, successfully applied to perceptual learning with explicit feedback, suggest that learning might be generalized to incorporate various feedback signals. However, it remains unclear whether the neural mechanisms for learning from explicit and implicit feedback are the same. This study directly compares these mechanisms using simultaneous EEG-fMRI during a perceptual decision-making task with intermixed explicit and no-feedback trials. The researchers hypothesize that learning from confidence utilizes distinct neural processes, possibly integrating with explicit feedback signals downstream to influence learning.

Extensive research demonstrates the impact of explicit feedback on learning within reinforcement learning frameworks. These models have been successfully applied to perceptual learning, where explicit feedback improves performance by shaping action selection, decision processes, and stimulus sensitivity. Conversely, studies show that perceptual learning occurs even without explicit feedback. The metacognitive evaluation of confidence plays a significant role in this context, acting as an internal feedback signal. While reinforcement learning models can incorporate different feedback signals, the neural mechanisms underlying learning from explicit versus implicit feedback remain largely unknown. This study builds upon previous work showing the involvement of the striatum and the dopaminergic system in reward processing, prediction error, and confidence, motivating the investigation into how these brain regions contribute to learning from different feedback sources.

Twenty-three participants performed a motion direction discrimination task (random dot kinematogram) during simultaneous EEG-fMRI acquisition. Task difficulty (coherence of dots) was adjusted across blocks to maintain performance (55-75% correct). Trials consisted of decision, bet (optional bet on accuracy), and feedback windows. Explicit feedback was provided on 50% of trials, while no-feedback trials required participants to infer the outcome based on confidence. EEG data were preprocessed to remove artifacts, downsampled, and filtered. Linear discriminant analysis (LDA) was employed to decode neural signatures of confidence (bet vs. no-bet) and explicit feedback (positive vs. negative) from the EEG. These decoded EEG signals were then used as regressors in a GLM analysis of the fMRI data, leveraging the trial-by-trial variability in the neural activity underlying subjective variables. A computational model comparison assessed which trials participants used confidence to learn from, comparing models that use confidence on only no-feedback trials, confidence on all trials and no confidence (learning only from explicit feedback). The behavioral data were assessed for learning by checking performance and sensitivity, and response perseveration (to check use of explicit feedback) . A psychophysiological interaction (PPI) analysis examined the functional connectivity between brain regions involved in integrating implicit and explicit feedback.

Behavioral data showed participants utilized confidence to inform their bets (higher accuracy and faster reaction times on bet trials), and learning from confidence on no-feedback trials. A computational model comparison demonstrated that a model which used confidence to learn on all trials best fit the behavioural data. EEG decoding revealed distinct neural signatures of implicit (confidence) and explicit feedback. In the decision window, the EEG confidence signal correlated with brain regions involved in decision-making and confidence computation. During the feedback window, this confidence-related EEG activity re-emerged, dissociating bet from no-bet trials in the absence of explicit feedback. A distinct EEG signature reflected the processing of explicit feedback. fMRI analysis revealed a dorsal-ventral striatal gradient, with implicit feedback representation more dorsal and explicit feedback more ventral. The external globus pallidus (GPe) showed a significant relationship with both confidence and explicit feedback signals. A Granger causality analysis suggested a connection between earlier BOLD activity in the inferior frontal gyrus (related to confidence) and later GPe BOLD. Finally, PPI analysis indicated that GPe BOLD modulated activity in the thalamus, insula, and rostromedial prefrontal cortex, influencing response perseveration, irrespective of feedback type.

The results support the hypothesis that confidence serves as implicit feedback for learning, even in the presence of frequent explicit feedback. The findings highlight the distinct neural encoding of implicit and explicit feedback values in the striatum, showing a dorsal-ventral gradient. The GPe's role as an integration hub for these feedback signals is supported by its correlation with both confidence and explicit feedback, and its influence on subsequent cortical activity. These results expand our understanding of how the brain integrates internal and external sources of information to guide learning and behavior. The findings add to the existing literature on the functional organization of the basal ganglia and its role in reinforcement learning and decision-making.

This study demonstrates the significant role of confidence as an implicit feedback signal in perceptual learning. The distinct neural encoding of implicit and explicit feedback values in the striatum, their integration in the GPe, and the downstream influence on cortical regions via the thalamus provide novel insights into the neural mechanisms underlying learning. Future research should investigate the precise computational mechanisms underlying the integration of implicit and explicit feedback in the GPe and explore the broader implications of this dual-feedback system for other cognitive domains.

The study design might be limited by the specific task used (motion direction discrimination) and the use of a simplified computational model. While the computational modelling supports the study's conclusions, a more complex model might be required to capture all aspects of the behavioural data. Also, the causality inferences are based on correlations and Granger causality which do not definitively show causal relationships. While the sample size was sufficient for detecting significant differences, a larger sample might improve the statistical power of the study.