CRII: CSR: Multi-View Learning Solutions for Next-Generation Computationally-Autonomous Wearables

可穿戴设备已经成为医疗保健、健身和以人为中心的物联网（IoT）领域许多新应用的革命性技术。包括机器学习和信号处理技术在内的计算算法经常用于从可穿戴传感器数据中连续实时地提取有价值的信息。然而，这些算法需要根据系统配置的任何变化进行重新训练，例如向网络中添加/移除传感器，传感器位移/错位，传感器升级，新用户采用系统，以及用户的物理和行为状态的变化。计算算法的再训练需要收集足够数量的标记训练数据，这是一个耗时、劳动密集型和昂贵的过程，限制了可穿戴技术的可扩展性和可持续性。本研究的目标是在不需要收集新的标记数据的情况下实现计算算法的自动重新配置。本研究旨在通过两个总体研究信托来设计、开发和验证可穿戴设备自配置的算法和工具。首先，本项目研究了源视图和目标视图同时观察感兴趣现象的场景的同步多视图学习解决方案。在同步学习中，源视图和目标视图的观察之间的直接关联是通过上下文敏感的学习过程建立的，该过程将生理监测和人体的特性考虑到迁移学习的目的。其次，本研究开发了异步多视图学习算法，即使在源视图和目标视图中没有同步测量，也可以实现知识的自动转移。异步学习研究设计了特征映射、实例转换和数据标记技术，以确定目标视图的数据实例如何与源视图的数据实例相关联，同时考虑到用户的物理和上下文属性。该项目可能会产生高度可持续和可扩展的可穿戴设备，能够在高度动态和不受控制的环境中进行自我监控和自我配置。计算自主可穿戴设备的真正实现将允许进行高精度的慢性疾病管理，并有助于新的基于可穿戴的消费者应用程序的可用性。这可能导致围绕以人为中心的物联网概念开发产品和业务，并在健康管理自动化和许多目前不可行的应用中使用它们。

Wearables have emerged as a revolutionary technology for many new applications in healthcare, fitness, and human-centered Internet-of-Things (IoT). Computational algorithms, including machine learning and signal processing techniques, are often used to extract valuable information from wearable sensor data continuously and in real-time. These algorithms, however, need to be retrained upon any changes in configuration of the system, such as addition/removal of a sensor to/from the network, sensor displacement/misplacement, sensor upgrade, adoption of the system by new users, and changes in physical and behavioral status of the user. Retraining of the computational algorithms requires collecting sufficient amount of labeled training data, a time consuming, labor-intensive, and expensive process that limits scalability and sustainability of wearable technologies. The goal of this research is to enable automatic reconfiguration of the computational algorithms without need for collecting new labeled data. This proposed research aims to design, develop and validate algorithms and tools for self-configuration of wearables through two overarching research trusts. First, this project investigates synchronous multi-view learning solutions for scenarios where source and target views observe the phenomena of interest simultaneously. In the synchronous learning, direct associations between observations made by the source view and those of the target view are established through context-sensitive learning processes that take the properties of physiological monitoring and human body into account for transfer learning purposes. Second, this research develops asynchronous multi-view learning algorithms to allow for automatic knowledge transfer even in absence of synchronous measurements in the source and target views. The asynchronous learning research devises feature mapping, instance transformation, and data labeling techniques to determine how data instances of the target view are associated with those of the source view while taking into consideration physical and contextual attributes of the user.This project will potentially result in highly sustainable and scalable wearables capable to self-monitor and self-configure in highly dynamic and uncontrolled environments. The true realization of computationally autonomous wearables will allow for conducting high-precision chronic disease management and contribute to availability of new wearable-based consumer applications. This can lead to the development of products and business around the concept of human-centered IoT and their use in automation of health management and many applications that are currently infeasible.