INSPIRE: Teaming Citizen Science with Machine Learning to Deepen LIGO's View of the Cosmos

该INSPIRE奖部分由计算机科学与工程局信息与智能系统部的网络人类系统计划，数学与物理科学局物理部的引力物理计划，和综合活动办公室。这个创新项目将开发一个公民科学系统，以支持先进激光干涉仪引力波aLIGO是引力物理学中有史以来最复杂的实验。在这个十年结束之前，它将打开宇宙引力波观测的窗口。然而，天体物理学发现所需的高探测器灵敏度使得aLIGO非常容易受到非宇宙伪像和噪声的影响，这些伪像和噪声必须从宇宙信号中识别和分离出来。教计算机识别和形态学分类这些文物的探测器数据是非常困难的。人类的视力是一种经过验证的分类工具，但来自大约30，000个传感器和监视器的aLIGO数据流很容易压倒一个人。这项研究将通过将人类分类与机器学习模型相结合来解决这些问题，该模型从公民科学家那里学习，并指导如何向参与者提供信息。该系统的一个新功能是依赖志愿者发现新的故障类别，而不仅仅是使用现有的故障类别。该项目包括对这个社会计算系统的以人为本的计算方面的研究，因此可以激发未来的公民科学项目，不仅利用志愿者的劳动，而且让他们成为科学发现的合作伙伴。 因此，该项目将为志愿者带来实质性的教育收益，他们将对科学如何运作有很好的理解，并将成为打开宇宙新窗口的兴奋的一部分。这是现有LIGO（当时正在进行技术改进）和Zooniverse（已经推出了可行的众包模式）之间的创新性跨学科合作，目前有超过一百万人参与了30个项目。 这项工作将帮助aLIGO快速识别科学数据流中的噪声和伪影，分离出合法的天体物理事件，并允许将这些事件分发到其他天文台进行更详细的源识别和研究。该项目还将建立和评估机器学习和人类学习之间的接口，这本身将是对当前方法的一种进步。它可以被描述为一个循环：（1）通过筛选大量的aLIGO数据，公民科学家将产生一个强大的“黄金标准”故障数据集，可用于种子和训练机器学习算法，以帮助识别任务。(2)将开发选择和分类故障事件的机器学习协议，以通过更有效的方式组织和传递数据，最大限度地发挥公民科学家的潜力。该项目将试验任务设计和工作流程组织（利用以前的Zooniverse经验），以构建一个系统，利用机器（系统处理大量数据的能力）和人类（识别模式和发现差异的能力）的独特优势，然后使用该模型实现高质量的aLIGO探测器表征和引力波搜索。

This INSPIRE award is partially funded by the Cyber-Human Systems Program in the Division of Information and Intelligent Systems in the Directorate for Computer Science and Engineering, the Gravitational Physics Program in the Division of Physics in the Directorate for Mathematical and Physical Sciences, and the Office of Integrative Activities.This innovative project will develop a citizen science system to support the Advanced Laser Interferometer Gravitational wave Observatory (aLIGO), the most complicated experiment ever undertaken in gravitational physics. Before the end of this decade it will open up the window of gravitational wave observations on the Universe. However, the high detector sensitivity needed for astrophysical discoveries makes aLIGO very susceptible to noncosmic artifacts and noise that must be identified and separated from cosmic signals. Teaching computers to identify and morphologically classify these artifacts in detector data is exceedingly difficult. Human eyesight is a proven tool for classification, but the aLIGO data streams from approximately 30,000 sensors and monitors easily overwhelm a single human. This research will address these problems by coupling human classification with a machine learning model that learns from the citizen scientists and also guides how information is provided to participants. A novel feature of this system will be its reliance on volunteers to discover new glitch classes, not just use existing ones. The project includes research on the human-centered computing aspects of this sociocomputational system, and thus can inspire future citizen science projects that do not merely exploit the labor of volunteers but engage them as partners in scientific discovery. Therefore, the project will have substantial educational benefits for the volunteers, who will gain a good understanding on how science works, and will be a part of the excitement of opening up a new window on the universe.This is an innovative, interdisciplinary collaboration between the existing LIGO, at the time it is being technically enhanced, and Zooniverse, which has fielded a workable crowdsourcing model, currently involving over a million people on 30 projects. The work will help aLIGO to quickly identify noise and artifacts in the science data stream, separating out legitimate astrophysical events, and allowing those events to be distributed to other observatories for more detailed source identification and study. This project will also build and evaluate an interface between machine learning and human learning that will itself be an advance on current methods. It can be depicted as a loop: (1) By sifting through enormous amounts of aLIGO data, the citizen scientists will produce a robust "gold standard" glitch dataset that can be used to seed and train machine learning algorithms that will aid in the identification task. (2) The machine learning protocols that select and classify glitch events will be developed to maximize the potential of the citizen scientists by organizing and passing the data to them in more effective ways. The project will experiment with the task design and workflow organization (leveraging previous Zooniverse experience) to build a system that takes advantage of the distinctive strengths of the machines (ability to process large amounts of data systematically) and the humans (ability to identify patterns and spot discrepancies), and then using the model to enable high quality aLIGO detector characterization and gravitational wave searches.