Learned Data Augmentation for Bias Correction

This thesis consists of three independent pieces of research that can be divided into two subject groups. The first block of topics is invariance learning and learned data augmentation (Paper 1 and 2 presented in Chapter 3 and 4, respectively). Paper 1 is concerned with learning invariances (or equivalently, as we will see, data augmentation) via Bayesian model selection and the marginal likelihood. In Paper 2, we take a different approach: achieving invariance by automatically pose-normalising inputs. The second topic block is fairness in machine learning which we cover in Paper 3 (Chapter 6). In addition to published research, this thesis contains the following original material. The first two chapters introduce the topics and Chapter 5 connects data augmentation with fairness. It investigates whether data augmentation and upsampling can be used make datasets more balanced, and, by correcting data bias, making models more fair. Chapter 7 concludes the work with a summary and discussion.