Neural Characteristic Activation Analysis and Geometric Parameterization for ReLU Networks

We show theoretically that GmP resolves the aforementioned instability issue. We report empirical results on various models and benchmarks to verify GmP's advantages of optimization