Demystifying Disagreement on-the-Line in High Dimensions
Faculty: Donghwan Lee, Behrad Moniri, Xinmeng Huang, Edgar Dobriban, and Hamed Hassani
Question: Evaluating the performance of machine learning models under distribution shift is hard when we only have unlabeled data from the target domain.
Contribution: We conducted the first comprehensive theoretical analysis of disagreement-on-the-line and discovered a variety of new phenomena that were not observed in the empirical studies.
*This work is accepted for publication at the International Conference on Machine Learning (ICML 2023)