Nonconvex learning algorithms are enabling transformative societal changes by revolutionizing how we process data. Despite wide empirical success, a satisfactory understanding of the behavior of these algorithms is still lacking. In particular, as systems and processes become increasingly automated with algorithms aiding or replacing human judgment, the importance of more reliable learning methodologies coupled with a thorough understanding of their behavior intensifies.
The overarching goal of my research is to develop guiding theory, scalable algorithms and tools that make the modern practice of nonconvex learning more principled, reliable and effective. I intend to contribute significantly to the transformation of modern data analysis from a collection of effective yet mysterious heuristics to a principled and completely reliable scientific discipline. Such a development will not only demystify existing practices but will facilitate new algorithms and system designs that better utilize computational and data resources.