Towards more resilient, robust, responsible decisions

Machine/deep learning work incredibly well... until it doesn't. In this presentation, I will present an approach producing resilient solutions (distributionnally robust optimization with Wasserstein uncertainty). I will emphasize the ideas, provide illustrations, and highlight the collaborative work on this topic at DAO. In particular, I will mention
(1) the statistical properties of robust models,
(2) a nice histogram reshaping,
(3) a toolbox (with scikit-learn and PyTorch interfaces) to robustify your own models !