My research explores the overlap between practical machine-learning methods and their
underlying theory - a space where recent progress in training deep networks still depends
largely on heuristics and
has little theoretical backing.
In my research, I aim to narrow the gap between theory and practice by
- developing practical theory that explains why specific phenomena arise and why certain heuristics succeed during deep-learning training,
- advancing existing algorithms by theoretically inspired modifications, and
- creating new methods with both strong convergence guarantees and practical performance.
