Seminar Details

Classical computational learning guarantees typically assume independent samples drawn from the same distribution on which a classifier will be evaluated. When the test distribution shifts or the training data is corrupted, however, these guarantees may no longer be trustworthy.

In this talk, I will present the first efficient algorithms for classification under distribution shift, where the learner must either certify low test error or detect harmful distribution shift. I will describe techniques based on sandwiching polynomials and iterative filtering that yield efficient guarantees for several fundamental concept classes.