CRCNS: Neural computations for continuous control in virtual reality foraging

点击翻译按钮获取中文摘要

Neuroscience has been able to gain major insights by relating measurements of neural activity to the brain’s sensory inputs and motor outputs. Yet most neural activity supports computations and cognitive functions (‘thoughts’) that are not directly measurable by the experimenter. The investigators for the present proposal invented a novel method to model an animal's thoughts by combining eXplainable Artificial Intelligence (XAI) cognitive models for naturalistic tasks with measurements of the animal’s sensory inputs and behavioral outputs. This model, called Inverse Rational Control (IRC), infers the internal model assumptions under which an animal's actions would be optimal. It then provides estimates of time series of subjective beliefs about the world that are consistent with this internal model. These estimates provide targets for a dimensionality reduction framework that assesses task-relevant computational dynamics within neural population activity. The investigators propose to use these analysis tools to find neural representations and transformations that implement these cognitive processes. They will apply this to a complex, naturalistic task that they developed: catching fireflies in virtual reality. The monkeys they successfully trained to perform this task demonstrably weigh uncertainty, develop predictions and long-term strategies, and apply nonlinear dynamics — all computations that are fundamental for brain function. The investigators propose first to apply their method to analyze existing behavioral data and neural recordings collected in a simple version of this task with a single target firefly. They will then collect new data on a multi-firefly version of the task, which incentivizes animals to make and implement longer-term plans. To analyze this data, the investigators will generalize their approach to allow them to learn which compressed representations are selected by the animal as the foundation for their strategies. These results will be used to form predictions about neural computations that will be tested using the electrophysiological data collected from multiple brain regions during this project. The results of this study will explain the computations required to perform a complex, strategic navigation task in the presence of uncertainty, and will demonstrate a new paradigm for understanding naturalistic brain computations. RELEVANCE (See instructions): This project will uncover the neural basis of cognitive processes in the primate brain that underlie spatial navigation, strategic planning, and behavioral control. It will demonstrate how a powerful new paradigm for understanding complex, natural brain computations can apply to a wide variety of tasks, to explain either adaptive or pathologically structured behavior. This will provide crucial guidance for understanding and improving disrupted human cognitive function.