Learning to Learn Across Domains

Abstract

In the previous chapters, we have seen several variants of cross-domain learning problems, including supervised and unsupervised domain adaptation, self-based learning for DA, and domain generalization. We have also seen modeling and algorithmic approaches that have been carefully designed to solve these problems using diverse ideas and techniques spanning from distribution alignment to style transfer and pseudo-labeling. Despite their diversity, many of these approaches share the commonality that they were designed by humans having prior intuitive or theoretical reason to believe that the introduced learning objectives or architectural changes would reduce the impact of the domain shift. By contrast, this chapter explores the emerging paradigm of learning-to-learn across domains. In this paradigm, the previously manual process of algorithm design and tuning is partially replaced or augmented by an additional learning process that trains some aspect of the algorithm or architecture to achieve empirically good performance in cross-domain learning.