CAREER: Probabilistic Models for Spatiotemporal Data with Applications to Dynamic Brain Connectivity

点击翻译按钮获取中文摘要

Probabilistic models are among the most promising tools for complex spatiotemporal data. However, transforming this promise to practical impact requires easy-to-deploy tools that appropriately address existing roadblocks. This project develops new tools for accurate and scalable probabilistic machine learning with spatiotemporal data. Furthermore, the approach is motivated by applications to mapping dynamic brain connectivity from human brain imaging data. The importance of dynamic brain connectivity lies in its description of neural information processing mechanisms, along with potentially transformative applications to understanding and treating neurological and neuropsychiatric disorders. This project will develop new techniques for estimating brain connectivity and apply these methods to the neuroscientific tasks of explaining inter-individual differences in cognition and behavior. This project will include curriculum development on probabilistic models for spatiotemporal data. This project also plans to involve participation by graduate students from underrepresented groups. This project creates a transformative new direction for modeling high-dimensional spatiotemporal data by addressing the fundamental challenges of modeling, scalability, and mitigating data biases. The first challenge is modeling, which refers to the inflexible assumptions of existing spatiotemporal models -- leading to under-fitting. To this end, this project develops modular probabilistic models that capture structured variability. Another pressing challenge is the computational scalability of inference and learning for such probabilistic models. This project tackles scalability by developing principled sample-selection methods for scalable approximate inference with performance guarantees. A third challenge is data bias, which occurs because data from a single source is often not statistically representative. Thus, models fit using single-source data have inconsistent and non-reproducible results. This project addresses data bias by combining data across multiple sources using novel federated learning for shared estimation without requiring direct data sharing. In addition to developing the algorithmic and theoretical frameworks for these directions, this project will also build and release open software.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.