33cc2b872dfe481abef0f61af181dfcf-AuthorFeedback.pdf

Hopefully though as we are approximating a12 convex loss this still greatly helps with the convergence and stability of training. It can be extended to non positive semidefinite matrices as in appendix eq. We have not investigated this rigorously, but anec-27 dotally this seems to be the case for large errors. We think this is strong evidence that our method is significantly better. We noticed some classes had rotation symmetries and we thought our method had interesting46 behaviourfortheseobjects.