CPS: Small: Human-in-the-Loop Learning of Complex Events in Uncontrolled Environments

该项目旨在推进人类在环网络物理系统的机器学习知识。移动和可穿戴设备已经成为一种很有前途的健康监测和行为干预技术。设计这样的系统需要通过主动学习过程在自由生活的环境中收集和标记传感器数据。在主动学习中，系统迭代地向人类专家（例如，患者，临床医生）查询正确的标签。在不受控制的环境中设计主动学习策略是具有挑战性的，因为(1)主动学习给用户带来了巨大的负担，并影响了技术的采用；(2)人类表达的标签具有显著的时间和空间差异，导致系统性能不佳。该研究将解决设计高性能系统的技术挑战，并在行为医学以外的许多应用中实现准确的监测和干预。该项目开发混合主动解决方案，通过以下研究目标，使在不受控制的环境中学习人类行为成为可能：(1)研究组合方法，以最大限度地提高主动学习性能，同时考虑到传感器数据的信息量、数据标记的负担和预期标签的可靠性；(2)基于知识图嵌入和半监督学习技术构建丰富的复杂行为词汇表；(3)开发基于网络图的学习算法，以推断复杂的人类行为；(4)通过室内实验和用户研究验证离线主动学习、实时主动学习、行为词汇构建和行为推理算法。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project aims to advance knowledge of machine learning for human-in-the-loop cyber-physical systems. Mobile and wearable devices have emerged as a promising technology for health monitoring and behavioral interventions. Designing such systems involves collecting and labeling sensor data in free-living environments through an active learning process. In active learning, the system iteratively queries a human expert (e.g., patient, clinician) for correct labels. Designing active learning strategies in uncontrolled settings is challenging because (1) active learning places a significant burden on the user and compromises adoption of the technology; and (2) labels expressed by humans carry significant amounts of temporal and spatial disparities that lead to poor performance of the system. The research will address technical challenges in designing high performance systems, and enable accurate monitoring and interventions in many applications beyond behavioral medicine.This project develops mixed-initiative solutions that will enable learning of human behaviors in uncontrolled environments through the following research objectives: (1) investigating combinatorial approaches to maximize the active learning performance taking into account informativeness of sensor data, burden of data labeling, and reliability of prospective labels; (2) constructing a rich vocabulary of complex behaviors based on knowledge graph embedding and semi-supervised learning techniques; (3) developing network-graph-based learning algorithms that infer complex human behaviors; and (4) validating algorithms for off-line active learning, real-time active learning, behavior vocabulary construction, and behavior inference through both in-lab experiments and user studies.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.