CRCNS: Inferring reference points from OFC population dynamics

点击翻译按钮获取中文摘要

A key computation that all mammals perform is determining the value of different outcomes. People and animal models evaluate outcomes as gains or losses relative to an internal reference point, likely reflecting their experience-based expectations. For example, if someone is told they will receive a particular salary at a new job, but when they start, they find that the salary is substantially less, they will view that salary (which is a net increase in wealth) as a loss relative to their reference point. Reference dependence is a consequential, ubiquitous phenomenon, driving decisions about insurance, financial products, labor, and retirement savings. The proposed work seeks to uncover how large populations of neurons represent a cognitive variable –the reference point- during value-based decision-making. This work involves complementary, synergistic interactions between experimentalists and theorists in the labs of Dr. Christine Constantinople and Dr. Cristina Savin, respectively. This proposal will develop a novel behavioral paradigm for studying reference dependence in rats, enabling application of powerful tools to monitor large-scale neural dynamics. High-throughput behavioral training will generate dozens of trained subjects for experiments in parallel. We will also develop a behavioral model to quantify key aspects of rats' behavior, including individual differences in behavior across animals (Aim 1). We will use new silicon probes with high channel counts (“Neuropixels” probes) to record from populations of neurons in dozens of rats during behavior. Recordings will be obtained from the orbitofrontal cortex (OFC), a key brain structure implicated in value-based decision-making. We will develop novel latent dynamics models that will infer the reference point directly from populations of simultaneously recorded neurons in OFC, without any knowledge of the task or rats' behavior. This model will also be able to identify aspects of neural dynamics that are common across dozens of rats, and aspects that are variable across animals, reflecting individual differences in behavior (Aim 2). Finally, we will use complementary, state-of-the-art machine-learning techniques to train recurrent neural networks (RNNs) on our behavioral and neural data. This approach will generate concrete hypotheses about the neural circuit architectures performing reference-dependent subjective valuation in our task (Aim 3).