Representational dynamics for flexible learning in complex environments

点击翻译按钮获取中文摘要

Project Abstract Humans display tremendous flexibility in their everyday behavior, adjusting it rapidly when appropriate (i.e. adopting mask wearing after onset of Covid-19), but not when inappropriate (i.e. continuing to drive after involvement in an unavoidable car accident). Recent work has highlighted the role that transient fluctuations in arousal, thought to be mediated by activation of the locus coeruleus norepinephrine (LC/NE) system, play in behavioral adjustments. Increasing NE pharmacologically promotes behavioral updating in rodents and peripheral measures of arousal, such as pupil diameter and P300 orienting response, provide a window into the dynamics that underlie these behavioral adjustments in humans. A mechanistic understanding of these processes could provide a valuable therapeutic target for a wide range of psychiatric disorders in which behavioral flexibility is impaired. However, current theory falls short, in part because it fails to account for the contextual nature of arousal: that heightened arousal reflects more behavioral adjustment in some settings or individuals, but less in others. We believe that previous computational accounts of NE have likely failed to explain heterogenous effects on behavior because they have ignored the neural representations on which NE acts. Recent advances in computational neuroscience have highlighted the importance of neural representations for efficient learning in complex environments, and provided tools to measure them. Building on this work, we developed a computational model in which NE drives transitions in neural representation that lead to behavioral adjustment when new representations persist in time (ie. after Covid), but reduce behavioral adjustment when they do not (after a freak accident). We propose that representational dynamics evoked by NE are not random, but instead are governed by assumptions about environmental structure, which differ across settings and individuals, to produce heterogeneous effects of arousal on behavior. This idea could facilitate personalized predictions for how NE manipulations would alter behavior, potentially enabling better treatment of attention deficit and anxiety disorders. Achieving this goal would first require basic research experiments to better characterize how and why arousal differentially relates to behavior across task contexts, individuals, and learning. Here we conduct these basic research studies, first measuring arousal by proxy in adversarial task structures (i.e. post-covid versus post-accident) to dissect the computational mechanisms through which it modulates behavior. Next, we examine internal representations directly, using fMRI in a task with ambiguous structure, to understand whether and how inter-individual differences in representational structure give rise to inter-individual differences in behavior and its sensitivity to arousal. Finally, we extend our existing computational model such that it can learn structure through experience and make individualized predictions for behavior across a wide range of possible environments. We test these predictions, as well as their relevance to various mental health constructs, in a large-scale online validation study.