A Discrepancy-Based Theory of Domain Adaptation

Watch a 54-minute lecture from the Simons Institute where Mehryar Mohri of Google Research and NYU presents a theoretical framework for domain adaptation centered on the concept of discrepancy. Explore both unsupervised and weakly supervised adaptation scenarios through detailed theoretical analysis of sample reweighting techniques. Learn about practical bounds that inform algorithm design for domain adaptation, particularly in scenarios with limited or no labeled target data. Examine experimental results demonstrating the effectiveness and robustness of these adaptation algorithms compared to baseline approaches.