Chapter 6 - Meta learning for domain generalization

Abstract

One of the fundamental issues of deep neural networks is their inability to generalize to diverse test environments, often yielding unreliable predictions in novel environments not seen during training. This sensitivity is undesirable, and can even be fatal in some safety-critical applications such as medical imaging. Domain generalization comprises of a class of techniques that aim to train generalizable models that can perform reliably in unseen test environments. In this chapter, we will look at some approaches that use meta learning for training domain-generalizable deep networks.