Learning the Geometry of Admissible Hypotheses through Inductive Bias in Training Distributions