III: Medium: Collaborative Research: Computational Tools for Extracting Individual, Dyadic, and Network Behavior from Remotely Sensed Data

III：媒介：协作研究：从遥感数据中提取个体、二元和网络行为的计算工具

• 批准号: 1514126
• 负责人: Brian Ziebart
• 金额: $ 55.43万
• 依托单位: University of Illinois at Chicago
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2015
• 资助国家: 美国
• 起止时间: 2015-09-01 至 2019-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1514126&HistoricalAwards=false
• 关键词: III; Medium; Collaborative; Research; Computational

Recent technological advances in location tracking, video and photo capture, accelerometers, and other mobile sensors provide massive amounts of low-level data on the behavior of animals and humans. Analysis of this data can teach us much about individual and group behavior, but analytical techniques that lead to insight about that behavior are still in their infancy. In particular, these new data can provide an unprecedented window into the lives of wild animals, augmenting the traditional time-consuming first-hand observations from field biologists. Unfortunately, the interpretation of low-level (i.e., unprocessed) data from animal-borne electronic sensors still poses a significant bottleneck in leveraging all of the available data to better understand the individual, pairwise, and group behavior of animal populations. This project will develop tools for scaling the expert knowledge needed to interpret high-level behaviors from low-level sensor data using tools from statistical machine learning and network analysis. These data and analytical tools promise to fundamentally change our understanding why animals do what they do, at high resolution and across multiple scales, from individuals to entire populations. The results of the project will be applicable in many settings where massive sensor data is overwhelming traditional insight derived from observational approaches. As part of the project, unique data on primate behavior that will bridge the low-level data and expert knowledge will be collected at Mpala Research Centre, Kenya. Undergraduate, graduate, and postdoctoral students from computer science and animal behavior will collaborate across continental and disciplinary boundaries. The technical aims of this project include developing structured prediction methods that improve behavior recognition at multiple levels (individual, pair-wise, and group), using network properties to improve the identification of group activities, and advancing active learning in the structured prediction setting so that "expensive" expert knowledge and supplemental data collection will be judiciously utilized for maximum benefit in learning behavior recognition models. Recognizing animal behavior from low-level sensor data is hierarchical in this approach, with individual activities recognized directly from data and the context of these data, the inferred individual activities informing pair-wise behavior recognition, and inferred pair-wise behavior informing group-level activity recognition. The benefits of improving the accuracy of individual and pair-wise behavior for recognizing group-level behavior will enable expert annotations to be requested that improve behavior recognition the most across all levels. These advances will enable field-biologists to investigate new hypotheses about fundamental evolutionary, ecological, and population processes at scale without the burdens of complete manual annotation of collected data. The methods will be applicable beyond field biology to understanding the hierarchy of behavior from individual entities to groups, from humans to cells, in scientific, educational, and business contexts. The team will leverage the interdisciplinary and international nature of the project to continue its ongoing work to increase participation of women and minorities in STEM research at undergraduate and graduate levels.

最近在位置跟踪、视频和照片捕捉、加速度计和其他移动传感器方面的技术进步提供了大量关于动物和人类行为的低级数据。对这些数据的分析可以教会我们很多关于个人和群体行为的知识，但能够洞察这些行为的分析技术仍处于起步阶段。特别是，这些新数据可以为野生动物的生活提供前所未有的窗口，增加了野外生物学家传统的耗时的第一手观察。不幸的是，对来自动物传播的电子传感器的低水平（即未经处理的）数据的解释仍然是利用所有可用数据来更好地理解动物种群的个体、成对和群体行为的重大瓶颈。该项目将开发工具，用于扩展所需的专家知识，使用统计机器学习和网络分析工具从低级传感器数据中解释高级行为。这些数据和分析工具有望从根本上改变我们对动物行为的理解，从高分辨率和跨多个尺度，从个体到整个种群。该项目的结果将适用于大量传感器数据压倒传统观测方法的许多环境。作为该项目的一部分，将在肯尼亚的Mpala研究中心收集有关灵长类动物行为的独特数据，这些数据将连接低级数据和专家知识。计算机科学和动物行为学的本科生、研究生和博士后将跨越大陆和学科界限进行合作。该项目的技术目标包括开发结构化预测方法，以提高多层次（个体，成对和群体）的行为识别，使用网络属性来改进群体活动的识别，并在结构化预测设置中推进主动学习，以便“昂贵”的专家知识和补充数据收集将被明智地利用，以获得学习行为识别模型的最大利益。在这种方法中，从低级传感器数据中识别动物行为是分层的，个体活动直接从数据和这些数据的上下文中识别出来，推断出的个体活动通知成对行为识别，推断出的成对行为通知群体活动识别。提高个体和成对行为识别群体行为的准确性的好处将使专家注释能够在所有级别上最大程度地提高行为识别。这些进步将使野外生物学家能够大规模地研究关于基本进化、生态和种群过程的新假设，而无需对收集到的数据进行完整的手工注释。这些方法将适用于野外生物学以外的领域，以理解从个体到群体、从人类到细胞的行为层次，以及在科学、教育和商业环境中的行为层次。该团队将利用该项目的跨学科和国际性质，继续其正在进行的工作，以增加女性和少数民族在本科和研究生阶段对STEM研究的参与。