Robust Domain Adaptation: Representations, Weights and Inductive Bias

Unsupervised Domain Adaptation (UDA) has attracted a lot of attention in the last ten years. The emergence Invariant Representations (IR) improved drastically transferability representations from labelled source domain to new and unlabelled target domain. However, potential pitfall this approach, namely presence label shift, been brought light. Some works address issue with relaxed version invariance obtained by weighting samples, strategy often referred as Importance Sampling. From our point view, theoretical aspects how Sampling interact UDA have not studied depth. In present work, we bound risk which incorporates both weights invariant representations. Our analysis highlights role inductive bias aligning distributions across domains. We illustrate it on standard benchmarks proposing learning procedure for UDA. observed empirically that weak makes adaptation more robust. elaboration stronger is promising direction algorithms.