Single-trial modeling separates multiple overlapping prediction errors during reward processing in human EEG.

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Reinforcement learning and cognitive control utilize valenced reward prediction errors (RPEs) and non-valenced salience prediction errors (PEs).
• A key debate in reward learning concerns the isolability of valenced and non-valenced PEs in human electroencephalogram (EEG) data.
• Existing research struggles to differentiate these distinct prediction error signals within complex neural activity.

Purpose of the Study:

• To disentangle sequential prediction errors (PEs) in the human EEG.
• To investigate whether valenced and non-valenced PEs can be distinctly identified.
• To reconcile debates on the neural representation of reward and salience PEs.

Main Methods:

• Combined behavioral modeling with single-trial EEG regression.
• Utilized an interval timing task that dissociates outcome valence, magnitude, and probability.
• Employed multiple regression across temporal, spatial, and frequency dimensions to analyze EEG data.

Main Results:

• Identified a spatio-tempo-spectral cascade of PEs, starting with early valenced RPE value and progressing to non-valenced RPE magnitude.
• Outcome probability was indexed by a late frontal positivity.
• Demonstrated that the apparent valenced RPE value effect is an artifact of overlapping non-valenced RPE magnitude responses (frontal theta feedback-related negativity for losses and posterior delta reward positivity for wins).

Conclusions:

• The study successfully separated valenced and non-valenced prediction error signals in human EEG.
• Findings suggest that distinct neural components previously attributed to valenced RPEs actually represent non-valenced magnitude signals.
• This work reconciles longstanding debates regarding the sequence and nature of neural components reflecting reward and salience PEs.