Goto

Collaborating Authors

 wideresnet160

Limitations Likeall lationisacoarse paths, andaggre intocoarse inpixelspace, hypothesized A.3 CIFAR-10 A.3.1 Label Figure 8 sho(Left)onall

A Appendix

Neural Information Processing Systems

Our implementations are based on Cubuk et al. The top-1 classification accuracies of these models are presented in Table 2. Model ImageNet Test Accuracy (%) Like all proxy measurements of spectral bias, linear interpolation is a coarse metric. A.3 CIF AR-10 full results A.3.1 Label Smoothing Figure 8 shows the same results as Figure 3 (Left) on all six CIFAR-10 models we tested. All six CIF AR-10 models we tested exhibit spectral bias. A.3.2 Sensitivity to Natural Image Directions Figure 9 shows the same results as Figure 5 on all six CIFAR-10 models we tested.