Robuta

https://openreview.net/forum?id=DbxKZvfOIhu Beyond BatchNorm: Towards a Unified Understanding of Normalization in Deep Learning | OpenReview We identify key properties in randomly initialized networks that accurately determine success and failure modes of different normalization layers https://openreview.net/forum?id=yL_iq-Q-ORS Perturbing BatchNorm and Only BatchNorm Benefits Sharpness-Aware Minimization | OpenReview Perturbing only the BatchNorm parameters in the adversarial step benefits several SAM-variants. sharpness aware minimizationbatchnormbenefitsopenreview