SoCS: Collaborative Research: A Human Computational Approach for Improving Data Quality in Citizen Science Projects

一个由计算机科学家、信息科学家、鸟类学家、项目经理和程序员组成的独特的跨学科团队将在机器学习方法和人类观察能力之间开发一个新的网络，以探索机械计算和人类计算之间的协同效应。这被称为人/计算机学习网络，虽然重点是在广泛的公民科学项目中提高数据质量，但该网络在各种复杂问题领域具有广泛适用性的潜力。该网络的核心是机器和人类之间的主动学习反馈环，它显著提高了两者的质量，从而不断提高网络的整体效率。人类/计算机学习网络将利用广泛征聘人类观察员的贡献，并用人工智能算法处理他们提供的数据，从而使总的计算能力远远超过其各自部分的总和。这项工作将以非常成功的eBird公民科学项目为试验平台，开发人/计算机学习网络。EBird雇佣了一个全球志愿者网络，他们每年向中央数据库提交数千万份鸟类观测数据。这项研究解决了公民科学中三个基本的数据质量挑战。它们是：1)减少识别或分类物体的错误；2)识别和量化个体观察者之间的差异；3)减少在许多公民科学项目中普遍存在的空间偏见。为了应对这些挑战，该项目将建立在人工智能进展的基础上，人工智能现在提供了通过生成能够解释巨大复杂性的模型来研究系统的机会。将通过开发新的多标签机器学习分类算法来扩展关于观察者分类的初步工作，以提供更好的生态解释和更准确的预测。此外，这项研究将通过构建抽样路径来开发新的主动学习算法，这些路径将优化志愿者调查工作，以最大限度地提高总体空间覆盖率，并通过众包技术激励参与。最后，它将研究参与者如何根据人工智能提供的反馈和信息来提高他们的观察质量。大规模的公民科学项目可以招募广泛的志愿者网络，他们在环境中充当智能和可训练的传感器，收集观察数据。人工智能过程可以显著提高志愿者可以提供的观测数据的质量，方法是根据观察员的专业知识过滤输入，这是一种基于汇总的历史数据的判断。通过向观察员提供关于观测精度的即时反馈和定制观测工作表，人工智能进程有助于提高观察员的专业知识，同时提高人工智能进程做出决定所依据的训练数据的质量。该项目的结果将对所有公民科学产生重大好处，并在无处不在的计算的新兴世界产生更广泛的影响，在这个世界中，人机合作将变得越来越普遍。

A unique interdisciplinary team of computer scientists, information scientists, ornithologists, project managers, and programmers will develop a novel network between machine learning methods and human observational capacity to explore the synergies between mechanical computation and human computation. This is called a Human/Computer Learning Network, and while the focus is to improve data quality in broad-scale citizen-science projects, the network has the potential for wide applicability in a variety of complex problem domains. The core of this network is an active learning feedback loop between machines and humans that dramatically improves the quality of both, and thereby continually improves the effectiveness of the network as a whole. The Human/Computer Learning Network will leverage the contributions of broad recruitment of human observers and process their contributed data with artificial intelligence algorithms leading to a total computational power far exceeding the sum of their individual parts. This work will use the highly successful eBird citizen-science project as a testbed to develop the Human/Computer Learning Network. eBird engages a global network of volunteers who submit tens of millions of bird observations annually to a central database.This research addresses three fundamental data quality challenges in citizen-science. These are: 1) reducing errors in identification or classification of objects; 2) identifying and quantifying the differences between individual observers; 3) reducing the spatial bias prevalent in many citizen-science projects. To address these challenges, the project will build on advances in artificial intelligence that now provide the opportunity to study systems through the generation of models that can account for enormous complexity. Preliminary work on observer classification will be extended by developing new multi-label machine learning classification algorithms that provide better ecological interpretations and more accurate predictions. In addition, the research will develop new active learning algorithms by constructing sampling paths that will optimize volunteer survey efforts to maximize overall spatial coverage, and incentivize participation via crowdsourcing techniques. Finally, it will study how participants can improve the quality of their observations based on the feedback and information provided by the artificial intelligence. Broad-scale citizen-science projects can recruit extensive networks of volunteers, who act as intelligent and trainable sensors in the environment to gather observations. Artificial intelligence processes can dramatically improve the quality of the observational data that volunteers can provide by filtering inputs based on observers' expertise, a judgment that is based on aggregated historical data. By guiding the observers with immediate feedback on observation accuracy and customization of observation worksheets, the artificial intelligence processes contribute to advancing expertise of the observers, while simultaneously improving the quality of the training data on which the artificial intelligence processes make their decisions. The results of the project will have significant benefit for all citizen science and broader impact in an emerging world of ubiquitous computing in which human-machine partnerships will become increasingly common.