2 ms·

If they’re optimizing over gradient descent, couldn’t they indirectly be doing some hessian-estimation methods?