Robustness to Noise in Machine Learning

To improve the trust we can pose in machine learning systems, we must dispose of robust algorithms for training new models and for testing existing ones. Such algorithms need to be theoretically sound and computationally efficient so that they can be used in practice. This project has the ambition of providing novel insights on the problem of learning from data affected by noise (of a statistical or even adversarial nature). I will achieve this by carrying out fundamental research to identify and tackle statistical and algorithmic challenges in robust training and testing.  The results of this project will have a twofold impact. First, they will improve the reliability of machine learning systems, which is crucial in safety-critical domains. Second, they will enlarge the scope where machine learning can be safely employed by relaxing the, typically idealized, assumptions that one has to place on the input data.