Non-Convex Optimization with Spectral Radius Regularization

Adam Sandler, Diego Klabjan, Yuan Luo

We develop regularization methods to find flat minima while training deep neural networks. These minima generalize better than sharp minima, yielding models outperforming baselines…