Learning with Average Top-k Loss