NeurIPS 2020

The Statistical Complexity of Early-Stopped Mirror Descent

Meta Review

The paper provides a clean and simple analysis of early-stopped mirror descent via a connection to offset Rademacher complexity. As noted by several reviewers, the downside is that linearity and square loss are essential for the proof to go through, but given the interest in the topic and approachability of the proofs, the paper will be of value to the broader NeurIPS community.