Brian Axelrod focuses on deep learning representations and loss geometry as a lever for understanding and precisely shaping model outputs.