The development of robust learning algorithms requires a deep understanding of the underlying data distribution.