Measures of relatedness in domain adaptation and related problems.

We will focus on unstructured distribution shifts, e.g., domain adaptation, transfer learning settings where little is assumed about the nature of the distribution shift. Of particular interest are discrepancy measures between source and target distributions that have been considered in the literature, how to adapt to unknown such measures, and implied statistical limits in various scenarios including multitask, and model selection. If time permits, we will discuss some new unifying principles that capture many such measures of discrepancy at once.