RTG: Advancing Machine Learning - Causality and Interpretability

RTG：推进机器学习 - 因果关系和可解释性

• 批准号: 1745640
• 负责人: Peng Ding
• 金额: $ 190.82万
• 依托单位: University of California-Berkeley
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2018
• 资助国家: 美国
• 起止时间: 2018-08-01 至 2024-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1745640&HistoricalAwards=false
• 关键词: RTG; Advancing; Machine; Learning; Causality

Faculty in the Statistics Department at the University of California, Berkeley develop an integrated program of research and education to support undergraduate research experiences, graduate research traineeships, and postdoctoral fellowships. The common research theme of the training activities is how to leverage the predictive power of statistical machine learning to address questions of causality and interpretability. The project aims to prepare the next generation of statisticians and data scientists to tackle new, important problems that arise from the analysis of massive data. Intuitively it seems that more reliable and precise inferences can be drawn from larger data sets. However, decisions and interventions must be interpretable and justified by statistical measures of uncertainty, which are challenging in this setting. This program will infuse ideas, energy, and resources in an integrated way at all levels of the educational program, from the undergraduate major to the postdoctoral experience, recruiting students and preparing them to participate in the extraordinary range of opportunities in this exciting new field.The research in this project will pursue theory to bridge the gap between causal inference and machine learning research, including high-dimensional inference, multiple testing, causal inference with interference, and causality and gene expression. The project is at the frontiers of statistics and data science, bridging the divide between machine learning and causal inference with potential impact far beyond the discipline of statistics. Plans are to redesign and expand the engagement of undergraduates in research through a graduate student mentorship program; to design new courses at the graduate and undergraduate levels, including an introductory course that builds on connections between data science, social sciences, and ethics; and to enhance graduate research training via a research symposium. The program will include a graduate professional development training series that addresses topics in technology, presentation and writing skills, and building an inclusive science community. The project will also provide significant training in teaching for graduate students and postdoctoral associates. Through a combination of channels, the innovations in training will spread to other institutions and disciplines, e.g., demonstrating the power of machine learning in policy and education settings where causal inference is central. The program also includes the development of educational materials with plans to disseminate them widely throughout the broader community. The project will emphasize recruitment and retention efforts targeted to increase the diversity of domestic students.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.

加州大学伯克利分校统计学系的教师开发了一个研究和教育的综合项目，以支持本科生的研究经历、研究生的研究培训和博士后奖学金。培训活动的共同研究主题是如何利用统计机器学习的预测能力来解决因果关系和可解释性问题。该项目旨在培养下一代统计学家和数据科学家，以解决大量数据分析中出现的新的重要问题。从直觉上看，似乎可以从更大的数据集中得出更可靠、更精确的推断。然而，决策和干预必须是可解释的，并通过不确定性的统计措施来证明其合理性，这在这种情况下是具有挑战性的。该项目将以一种整合的方式为教育项目的各个层面注入思想、精力和资源，从本科专业到博士后经验，招募学生并为他们参与这个令人兴奋的新领域的非凡机会做好准备。本项目的研究将通过理论来弥合因果推理和机器学习研究之间的差距，包括高维推理、多重测试、带干扰的因果推理、因果关系和基因表达。该项目处于统计学和数据科学的前沿，弥合了机器学习和因果推理之间的鸿沟，其潜在影响远远超出了统计学的范畴。计划是通过研究生指导计划重新设计和扩大本科生对研究的参与；在研究生和本科生阶段设计新课程，包括建立在数据科学、社会科学和伦理学之间联系的入门课程；并通过研究研讨会加强研究生的研究训练。该项目将包括一个研究生专业发展培训系列，涉及技术、演讲和写作技巧等主题，以及建立一个包容性的科学社区。该项目还将为研究生和博士后提供重要的教学培训。通过多种渠道的结合，培训方面的创新将传播到其他机构和学科，例如，在因果推理为中心的政策和教育环境中展示机器学习的力量。该方案还包括编写教育材料，并计划在更广泛的社区中广泛传播。该项目将强调以增加国内学生多样性为目标的招生和留用工作。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。

项目成果

Synthetic controls with staggered adoption

• DOI: 10.1111/rssb.12448
• 发表时间: 2021-12-07
• 期刊: JOURNAL OF THE ROYAL STATISTICAL SOCIETY SERIES B-STATISTICAL METHODOLOGY
• 影响因子: 5.8
• 作者: Ben-Michael, Eli;Feller, Avi;Rothstein, Jesse
• 通讯作者: Rothstein, Jesse

The Three Stages of Learning Dynamics in High-dimensional Kernel Methods

高维核方法中学习动力学的三个阶段

• 发表时间: 2021
• 期刊: ArXivorg
• 作者: Nikhil Ghosh, Song Mei

More Style, Less Work: Card-style Data Decrease Risk-limiting Audit Sample Sizes

更多风格，更少工作：卡片式数据减少风险限制审计样本量

• DOI: 10.1145/3457907
• 发表时间: 2021
• 期刊: Digital Threats: Research and Practice
• 作者: Glazer, Amanda K.;Spertus, Jacob V.;Stark, Philip B.
• 通讯作者: Stark, Philip B.

Deconstructing Distributions: A Pointwise Framework of Learning

• 发表时间: 2022-02
• 期刊: ArXiv
• 作者: Gal Kaplun;Nikhil Ghosh;S. Garg;B. Barak;Preetum Nakkiran

Reading to write

读来写

• DOI: 10.1111/1740-9713.01469
• 发表时间: 2020
• 期刊: Significance
• 作者: Nolan, Deborah;Stoudt, Sara
• 通讯作者: Stoudt, Sara