NeurIPS 2022, "Spherization Layer: Representation Using Only Angles"

Spherization layer is an explicit solution for the dispersion to completely eliminate the interference of the Euclidean norms in training without drawbacks