Knowledge Distillation as Semiparametric Inference [Talk]

More accurate machine learning models often demand more computation and memory at test time, making them difficult to deploy on CPU- or memory-constrained devices. Knowledge distillation alleviates this burden by training a less expensive student model to mimic the expensive teacher model while maintaining most of the original accuracy. To explain and enhance this phenomenon, we cast knowledge distillation as a semiparametric inference problem with the optimal student model as the target, the unknown Bayes class probabilities as nuisance, and the teacher probabilities as a plug-in nuisance estimate. By adapting modern semiparametric tools, we derive new guarantees for the prediction error of standard distillation and develop two enhancements—cross-fitting and loss correction—to mitigate the impact of teacher overfitting and underfitting on student performance. We validate our findings empirically on both tabular and image data and observe consistent improvements from our knowledge distillation enhancements.

Learn more about the 2020-2021 Directions in ML: AutoML and Automating Algorithms virtual speaker series >

Speaker Bios

Lester is a statistical machine learning researcher at Microsoft Research New England and an adjunct professor at Stanford University. He received his Ph.D. in Computer Science (2012), his M.A. in Statistics (2011) from UC Berkeley, and his B.S.E. in Computer Science (2007) from Princeton University. Before joining Microsoft, Lester spent three wonderful years as an assistant professor of Statistics and, by courtesy, Computer Science at Stanford and one as a Simons Math+X postdoctoral fellow, working with Emmanuel Candes. Lester’s Ph.D. advisor was Mike Jordan, and his undergraduate research advisors were Maria Klawe and David Walker. He got his first taste of research at the Research Science Institute and learned to think deeply of simple things at the Ross Program. Lester’s current research interests include statistical machine learning, scalable algorithms, high-dimensional statistics, approximate inference, and probability. Lately, he’s been developing and analyzing scalable learning algorithms for healthcare, climate forecasting, approximate posterior inference, high-energy physics, recommender systems, and the social good.