Learning and Decision Making with Preferential Supervision
Despite the widespread use and success of machine learning, a key bottleneck in applying machine learning is the need for high quality supervision that can be used to guide training of the algorithms. However, obtaining meaningful labels and designing rewards for supervision is becomes challenging, particularly as machine learning is used to solve increasingly complex problems. Poorly designed rewards and inaccurate labels can result in unstable and unsafe performance. This necessitates use of alternative forms of supervision for learning and decision making. In the setting of human-in-the-loop, preferences in the form of pairwise comparisons or rankings have emerged as an alternate supervision mechanism that are often easier to elicit and more accurate than labels or rewards. This talk will outline our efforts in understanding the fundamental limits of learning and decision making when an algorithm is given access to preferences in addition to labels. We will discuss and contrast the value of preferential supervision in several settings including classification, regression, bandits, optimization and reinforcement learning, along with some open problems.
Part of the Data Science Institute Distinguished Speaker Series:
Defining The Field of Data Science
As data science evolves from buzzword to a mature and singular field, its research questions dive deeper into the foundations of this new discipline. The Fall 2021 Distinguished Speaker Series convenes world-class experts actively exploring and expanding the fundamental methods and approaches that transform large and complex datasets into knowledge and action, fueling new applications in areas such as artificial intelligence, healthcare, and the social sciences. Join the new UChicago Data Science Institute for provocative talks and discussion that will illuminate the bedrock and promise of the flourishing field of data science.
Host: Data Science Institute
Aarti Singh is an Associate Professor in the Machine Learning Department within the School of Computer Science at Carnegie Mellon University. She received her Ph.D. degree in Electrical and Computer Engineering from the University of Wisconsin-Madison and was a Postdoctoral Research Associate at the Program in Applied and Computational Mathematics at Princeton University before joining CMU. Her research lies at the intersection of machine learning, statistics and signal processing, and focuses on developing, analyzing and applying interactive algorithms that use the most informative data and actions to guide learning and decision-making in both human-in-loop and human-out-of-loop settings, with applications to enabling social and scientific discoveries. Her work is recognized by an NSF Career Award, a United States Air Force Young Investigator Award, A. Nico Habermann Faculty Chair Award, Harold A. Peterson Best Dissertation Award, and four best student paper awards. Dr. Singh has served as Program Chair for the International Conference on Machine Learning (ICML) 2020, Program Chair for Artificial Intelligence and Statistics (AISTATS) 2017 conference, the National Academy of Sciences (NAS) committee on Applied and Theoretical Statistics, lead expert on ONR/NIST and NAS studies, NASEM advisory board for NSF DMREF, and Associate Editor for IEEE Transactions on Information Theory.