Towards Problem-dependent Optimal Learning Rates