III: Medium: Collaborative Research: Principled Uncertainty Quantification in Deep Learning Models for Time Series Analysis

点击翻译按钮获取中文摘要

Time series data are ubiquitous in modern science and engineering. An unprecedented amount is being collected in diverse applications such as healthcare systems, the Web, cyber network monitoring, self-driving cars, and Internet-of-Things services. While deep learning has achieved enormous success in time series predictive analysis, a key bottleneck of such models is that they are ignorant about the uncertainties in their predictions. A consequence is that they can produce wildly wrong predictions without noticing---this will lead to misguided decisions, which can be catastrophic in life-critical applications. This project aims to remedy this issue and advance deep learning towards more trustworthy time series analysis. The project will enable principled deep learning models for uncertainty-aware and reliable time series regression and classification without sacrificing their predictive power. Research findings from the project will be incorporated into graduate-level classes, tutorials, and workshops to bring multiple stakeholders and domain scientists together.The technical aims of this project are divided into three thrusts. First, the project will develop novel techniques bridging deep sequential models (e.g., recurrent networks, transformers) with Gaussian processes to quantify uncertainty in the functional space. Second, the project will explore how to learn calibrated deep sequential models and how to further decouple different sources of uncertainties to understand where a model's predictive uncertainty comes from. Third, the project will harness uncertainty to improve the reliability and efficiency of time series predictive systems. These techniques will enjoy the representation power of deep neural networks for modeling complex temporal dependencies in time-series data, while providing principled methodologies for quantifying and leveraging uncertainty for robustness and performance. The developed new models, algorithms, and techniques will be deployed in two important applications for times series analysis: 1) public health monitoring and forecasting, and 2) real-time analysis for mobile sensing time series data. The developed tools will also be open-sourced for trustworthy time series analysis that can benefit many other applications.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.