Appendices

The appendix is organized as follows. In Appendix B, we prove that CE, FL and LS satisfy the contrastive propertyinDefinition1. Let p,q be positive real numbers satisfying 1p + 1q = 1. The figure depicts the NC1 and test accuracy of various iteration-width configurations for different loss functions onCIFAR10. In this section, we first describe more details about the datasets and the computational resource used in the paper.