Feedback Discounting in Probabilistic Categorization: Converging Evidence from EEG and Cognitive Modeling

In simple probabilistic learning environments, the informational value of corrective feedback gradually declines over time. This is because prediction errors persist despite learners acquiring the contingencies between stimuli and outcomes. An adaptive solution to the problem of unavoidable prediction error is to discount feedback from the learning environment. We provide novel neural evidence of feedback discounting using a combination of behavioral modeling and electroencephalography (EEG). Participants completed a probabilistic categorization task while EEG activity was recorded. We used a model-based analysis of choice behavior to identify individuals that did and did not discount feedback. We then contrasted changes in the feedback-related negativity (FRN) for these two groups. For individuals who did not discount feedback, we observed learning-related reductions in the FRN that reflected incremental changes in choice behavior. By contrast, for individuals who discounted feedback, we found that the FRN was effectively eliminated due to the rapid onset of feedback discounting. The use of a feedback discounting strategy was linked to superior performance on the task, highlighting the adaptive nature of discounting when trial-to-trial outcomes are variable, but the long-term contingencies relating cues and outcomes are stable.