CAREER: Robust, scalable, reliable machine learning

机器学习越来越多地应用于大规模的关键任务问题中，目的是做出影响大量个人就业、储蓄、健康和安全的决策。机器学习有可能极大地影响和改变人们的生活，这就要求机器学习方法必须是健壮的、可解释的、可理解的，而不是黑盒子。这项研究开发了新的技术，这些技术既具有计算动机，又在理论上适用于大规模的强大机器学习。这项工作是在三种现代应用程序的背景下进行的。经济学家对分析小额信贷的有效性很感兴趣，小额信贷是指以消除贫困为目标向贫困地区的个人提供的小额贷款。(2)生物学家对利用单细胞RNA测序数据来了解细胞之间的关系和发育轨迹感兴趣。(3)物联网（IoT）有望产生大量复杂的数据，包括建筑物、交通基础设施、道路上的车辆以及许多其他传感器来源的能源读数。PI直接与领域专家合作，以便在应用程序领域产生直接、广泛的影响。在该项目的教育部分，PI是开发麻省理工学院统计、数据科学和统计机器学习的新研究生课程和学位的核心部分。本提案中的方法和应用是现代机器学习方法新课程的特色。PI还在开发高中水平的机器学习入门课程，作为已建立的女性技术计划（WTP）的一部分。鲁棒性和可解释性的问题尤其出现在具有重要空间和时间依赖性的领域，这些领域的数据量通常是巨大的，并且从业者在使用特定数据集之前通常对该领域有一些专业知识。这些正是现有机器学习方法发展不够完善的领域。需要将结构知识用于解决问题，建议使用贝叶斯方法，该方法可以通过先验和建模假设合并这些知识。然而，为了实现这些方法的承诺，实际方法需要对假设以及嘈杂或对抗性数据具有鲁棒性，以免这些数据以从业者无法理解的方式改变重要决策。这项研究结合了统计物理学的进展，以评估数据分析对假设和数据值的敏感性。为了实现所提出的健壮且可理解的机器学习框架的优势，从业者必须面对极端的可扩展性问题——无论是从计算角度还是从建模角度。在计算方面，这项研究建立在从计算几何到规模到现代规模的数据集的最新进展之上。在建模方面，请注意，虽然小规模问题表现出密集的时空依赖关系，但大规模问题往往更稀疏，并且实际方法必须反映这种稀疏性才能在规模上可靠。这项工作结合了概率论的进步来模拟稀疏物联网网络。该提案是高度跨学科的——汇集了机器学习、统计学、物理学、理论计算机科学、概率论和系统的想法，并将这些想法应用于小额信贷、单细胞RNA测序、传感器网络、国际贸易和工业应用，包括大规模的客户服务。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。

Machine learning is increasingly deployed in large-scale, mission critical problems for the purpose of making decisions that affect a vast number of individuals' employment, savings, health, and safety. The potential for machine learning to dramatically impact and change people's lives necessitates that machine learning methods be robust, explainable, and understandable---rather than black-box. This research develops new techniques that are both computationally motivated and theoretically sound for robust machine learning at scale. The work is situated in the context of three modern classes of applications. (1) Economists are interested in analyzing the efficacy of microcredit, small loans to individuals in impoverished areas with the goal of eliminating poverty. (2) Biologists are interested in using single-cell RNA sequencing data to understand cells' relationships and development trajectories. (3) The Internet of Things (IoT) is poised to generate a wealth of complex data across energy readings in buildings, within transportation infrastructure, from vehicles on the road, and from many other sensor sources. The PI is working directly with area experts so as to have immediate, broad impact across application domains. In an educational component of the project, the PI is a core part of developing a new graduate curriculum and degree in statistics, data science, and statistical machine learning at MIT. The methods and applications in this proposal feature in a new course on modern machine learning methods. The PI is also developing a high-school level introduction to machine learning as part of the established Women's Technology Program (WTP).The issues of robustness and explainability particularly arise in domains with nontrivial spatial and temporal dependencies, where the amount of data is often massive, and where practitioners typically have some expert knowledge about the domain before engaging with a particular dataset. These are precisely the domains where existing machine learning methodologies are less well-developed. The need to bring structural knowledge to bear on the problem suggests the use of Bayesian methods, which can incorporate this knowledge via prior and modeling assumptions. To live up to the promise of these methods, though, practical approaches need to be robust to assumptions as well as to noisy or adversarial data, lest this data change important decisions in ways not understood by the practitioner. This research incorporates advances in statistical physics to assess the sensitivity of a data analysis to assumptions and data values. And to realize the advantages of the proposed robust and understandable machine learning framework, practitioners must face extreme scalability issues---both from a computational perspective as well as a modeling perspective. On the computational side, this research builds on recent advances from computational geometry to scale to data sets at modern sizes. On the modeling side, note that while small-scale problems exhibit dense spatio-temporal dependencies, large-scale problems tend to be sparser, and practical approaches must reflect this sparsity to be reliable at scale. This work incorporates advances in probability theory to model sparse IoT networks. This proposal is highly interdisciplinary---bringing together ideas from machine learning, statistics, physics, theoretical computer science, probability theory, and systems and applying these ideas to microcredit, single-cell RNA sequencing, sensor networks, international trade, and industrial applications including customer service at scale.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.