Electrophysiological Signatures of Reward, Surprise, and Conflict in the Human Brain
Flexible, intelligent behavior requires choosing actions predicted to lead to optimal outcomes in an ever-changing environment and learning from feedback to improve future predictions. This action-feedback loop is governed by cognitive control, which allocates attention, working memory, and decision making resources according to changes in task demands. Predictive coding frameworks describe these dynamics in terms of prediction errors (PEs), or the difference between expected and actual outcomes. Despite decades of cognitive neuroscience research, the neural mechanisms of cognitive control remain elusive, in part because complex cognition depends on rapid interactions between widespread brain networks. Studying cognitive control in humans provides the opportunity to design complex tasks that dissociate the various computations underlying rich behaviors, but these advantages are typically offset by the limited spatial or temporal resolution of non-invasive measures of neuronal activity. This dissertation pairs behavioral modeling with the combined spatial and temporal resolution of direct brain recordings in humans to test foundational theories of cognitive control and predictive coding. Chapter 2 addresses a longstanding debate over whether scalp electroencephalography (EEG) signatures represent a valenced, quantitative reward prediction error (RPE) or the non-valenced magnitude of RPE. We use reinforcement learning principles to model individual participant behavior in an interval timing task and apply powerful single-trial regression analyses across time, space, and frequency dimensions to disentangle multiple overlapping components. Our results show valenced RPE effects are an artifact of component overlap between the early, frontal, theta frequency feedback-related negativity tracking non- valenced RPE magnitude for negative outcomes and the subsequent, more posterior reward positivity in delta frequencies tracking non-valenced RPE magnitude for positive outcomes. Our modeling approach also uncovered a novel, late frontal P300-like component elicited by low probability outcomes. Chapter 3 uses the same paradigm and behavioral model to investigate RPE value and magnitude coding at the local circuit level based on high frequency broadband (HFB) activity extracted from intracranial EEG (iEEG) recording in lateral prefrontal cortex (LPFC), medial prefrontal cortex (MPFC), and insular cortex (IC). We show that many sites in each of these three control regions represent either RPE value or magnitude, but also that some electrodes show mixed selectivity for both RPE features. Interestingly, RPE value and magnitude representations were most common in IC, which also showed the greatest proportion of electrodes with mixed selectivity. Furthermore, onsets of RPE value effects were earlier in IC than MPFC, suggesting a potential leading role for IC in reward processing. Finally, Chapter 4 characterizes the spatiotemporal evolution of conflict signals using HFB activity in a color-word Stroop task. In order to distinguish detection, resolution, and monitoring phases of conflict processing, we use a dynamic sliding window analysis strategy to segregate effects into stimulus, decision, and responses stages of the trial. We observed widespread conflict effects in LPFC, MPFC, IC, orbitofrontal, sensorimotor, and temporal cortices that formed partially overlapping but largely distinct networks for stimulus, decision, and response stages. Contrary to serial processing hypotheses proposed by classic conflict monitoring theory, we found the onsets of these conflict signals to be heterogeneous within and across regions. These results indicate conflict processing unfolds across distributed networks that work in parallel throughout the trial. Future studies can extend these findings using connectivity network analyses to determine information flow between these networks, which will help constrain the functional roles of each region. In particular, analyzing the relationship between HFB activity and rhythmic low frequency responses will help identify circuit mechanisms that bridge local HFB activity and non-invasive biomarkers from scalp EEG. Collectively, the experiments in this thesis provide novel insights into how reward, surprise, and conflict signals are processed in parallel across distributed networks and emphasize the importance of combining experimental design, behavioral modeling, and advanced signal processing to understand the neural computations underlying cognitive control in the human brain.