Affective arousal as a guide for reinforcement learning in multidimensional environments

点击翻译按钮获取中文摘要

Project Summary/Abstract In 2020, affective disorders are expected to impact over 35 million U.S. adults, imposing an economic burden in excess of $200 billion, and exacting an immeasurable personal toll on those affected. To mitigate these costs, the fields of neuroscience and psychology have devoted considerable resources to basic research aimed at identifying vulnerabilities, treatments, and prevention strategies for affective disorders. These investments have been slow to translate into concrete improvements, however, due in part to the enormous explanatory gap between neural and mentalistic accounts of affect. Faster progress requires a bridge from measurements at the biological level to phenomena at the psychological level. With a three-year NRSA fellowship, I will begin to develop this bridge by studying affect within the framework of computational reinforcement learning (RL), which has emerged as an indispensable guide for linking neural activity to psychological phenomena. I propose to use innovative RL models to bridge behavioral and biological data from two neuroimaging studies, both of which focus on affect and its role in attention and learning. These studies address a deep psychological problem: At any moment, the options for what a brain can attend to and learn about are infinite, so how do brains decide what to focus on and what to ignore? Consider, for example a seemingly simple task commonly encountered in neuroscience experiments: learning associations between words and pictures of scenes. Scenes vary along countless dimensions (location, habitability, beauty, etc.), so how do brains decide which dimensions to associate with the words, and which to ignore? I propose that such decisions are guided by affective arousal. Specifically, I hypothesize that high arousal directs attention towards dimensions that can take on a small number of values (e.g., whether the scene is indoors or outdoors, a dimension that can take on just two values). I call these low-cardinality dimensions, in contrast to high- cardinality dimensions, which can take on many values (e.g., the specific location of the scene). My hypothesis builds on research showing that brains in states of high arousal opt for fast, efficient learning strategies; all else being equal, low- cardinality dimensions are relatively easy to learn about, and high-cardinality dimensions are relatively difficult to learn about, so high arousal should direct attention towards the former and away from the latter. Testing this hypothesis will help launch my career at the intersection of psychology and computational neuroscience — a transdisciplinary approach that, I believe, will play an essential role in translating basic research into clinical applications. Specifically, I will develop the computational and neuroscience skills necessary to bridge the explanatory gap between neural and mentalistic accounts of affect, thereby aiding in the understanding, prediction, and treatment of affective disorders.