7990ec44fcf3d7a0e5a2add28362213c-AuthorFeedback.pdf

It is true that the decoding overB goes a long way, since it makes sure that a valid permutation matrix is always44 predicted. But we argue that Proposition 1 is insightful: even using[0,1]k instead Rk improves test accuracy a lot,45 according toourexperiments.