Bayes without Underfitting: Fully Correlated Deep Learning Posteriors via Alternating Projections

Open in new window