Uniform Priors for Meta-Learning