New Mathematical Approaches to Interpretability and Robustness that Directly Confront the High Dimensionality and Nonlinearity of Neural Representations