Should We Learn Most Likely Functions or Parameters? Shikai Qiu Tim G. J. Rudner Sanyam Kapoor Andrew Gordon Wilson New York University