CAREER: Foundations of Human-Centered Machine Learning in the Wild

今天的机器学习模型必须在日益动态和开放的环境中运行。开放环境的一个重要特征是，智能系统将遇到新的上下文和分布外的数据;也就是说，原始系统中没有用于训练算法的数据。然而，目前的监督学习是脆弱的，缺乏对数据随时间演变的方式以及如何应对不断变化的环境的可靠理解。这带来了一系列新的挑战，并推动了重新思考机器学习算法设计的需求，这些算法可以检测分布外的数据，并在野外不断变化的数据下修复学习模型。该项目将直接影响许多现实世界的领域，包括自主交通，医疗保健，商业和科学发现。该项目的目标是为开放世界环境中的安全，自适应和长期有益的学习算法奠定新的基础。该项目有三个连续的研究目标，将从根本上改变机器学习模型在野外训练、更新和监控的方式：（1）创建一个新的学习框架，以实现可靠的决策，在野外部署机器学习模型时，对未知情况提供强大的安全性;（2）加速模型适应，学习对野外出现的新概念进行分类，同时最大限度地减少所需的人类监督;（3）从长期准确性和安全性的角度描述和理解动态，最大限度地发挥模型在长期发展和运行中的影响。该教育计划将通过一门新课程、一个新的本科生导师计划“进入人工智能研究”和外展工作来宣传开放世界机器学习的力量。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Machine learning models today must operate amid increasingly dynamic and open environments. One important characteristic of open environments is that the intelligent system will encounter new contexts and out-of-distribution data; that is, data that were not used to train the algorithms the original system. However, current supervised learning is brittle, lacking a reliable understanding of the way data evolves over time and how to respond to changing environments. This introduces a set of new challenges and drives the need for rethinking the design of machine learning algorithms, that can detect out-of-distribution data and repair the learned model under evolving data in the wild. The project will directly impact many real-world domains including autonomous transportation, healthcare, commerce, and scientific discovery.The objective of this project is to lay new foundations for safe, adaptive, and long-term beneficial learning algorithms in open-world environments. The project has a continuum of three research aims that will fundamentally transform the way that machine learning models are trained, updated, and monitored in the wild: (1) create a new learning framework for reliable decisions, rendering strong safety against unknowns upon deploying machine learning models in the wild; (2) accelerate model adaptation, learning to classify new concepts emerging in the wild while minimizing human supervision required; (3) characterize and understand dynamics in terms of long-term accuracy and safety, maximizing the impact of models as they evolve and operate in the long run. The education plan will publicize the power of open-world machine learning through a new course, a new undergraduate mentorship program 'Entering AI Research' and outreach efforts.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.