7 Penn Papers in Top Machine Learning Conference!

The Neural Information Processing Conference, NeurIPS, formerly known as NIPS, is one of the leading venues for machine learning researchers.  Penn faculty, students (indicated by an * below), and alumni (indicated by a + below) were very well represented!  Aaron Roth and Shivani Agarwal served as Area Chairs, and we had no fewer than 7 papers:

• Local Differential Privacy for Evolving Data, by Matthew Joseph*, Aaron Roth*, Bo Waggoner+, and Jonathan Ullman.  This Spotlight paper studies how to maintain a statistic with local differential privacy with privacy loss that degrades only with the number of times the statistic changes, not with the number of times you update it.
• A Smoothed Analysis of the Greedy Algorithm for Linear Contextual Bandits, by Sampath Kannan*, Jamie Morganstern+, Aaron Roth*, Bo Waggoner+, and Steven Wu+.  This Spotlight paper explains why generically you don’t need exploration in linear contextual bandit problems: greedy is enough, so long as the contexts are slightly perturbed.
• Online Learning with an Unknown Fairness Metric, by Stephen Gillen*, Christopher Jung*, Michael Kearns*, and Aaron Roth*.  This paper shows how to solve an online learning problem subject to an individual fairness metric constraint, when you don’t know the metric, but are friends with a moral philosopher.
• Learning Loop Invariants for Program Verification, by Xujie Si*, Hanjun Dai, Mukund Raghothaman*, Mayur Naik*, and Le Song.  This Spotlight paper shows how to infer loop invariants, a fundamental problem in software verification, using deep reinforcement learning.
• Verifiable Reinforcement Learning via Policy Extraction, by Osbert Bastani*, Yewen Pu, and Armando Solar-Lezama. We show how to learn control policies for robotics tasks that we can automatically prove are safe; e.g., they avoid obstacles, do not fall over, etc.
• Contextual Pricing for Lipschitz Buyers, by Jieming Mao*, Renato Paes Leme, and Jon Schneider.  This paper shows how to maximize profit when many different kinds of goods to buyers with unknown valuations.
• Learning Pipelines with Limited Data and Domain Knowledge: A Study in Parsing Physics Problems, by Mrinmaya Sachan, Avinava Dubey, Tom Mitchell, Dan Roth*, and Eric P. Xing. This paper studies joint learning and inference for parsing (and answering) Newtonian Physics problems in textbooks.

Congratulations to all of the authors!