Alexander G. Gray

We demonstrate scenarios where sparse attention based transformer models learn and generalize faster, and theoretically characterize conditions under which this occurs.