Estimating the expected output of wide random MLPs more efficiently than sampling

Open in new window