Appendices A Proofs Proposition 4

F or the cross-entropy loss, Eq. (6) induces noise For Eq. (6), the noisy gradient is null We use the dataset as the open-set auxiliary dataset for experiments with CIFAR-10 and CIFAR-100. Gaussian noises have each dimension i.i.d. Blobs noises consist in algorithmically generated amorphous shapes with definite edges. There are ten classes comprised of the digits 0-9. Figure 1c, we use 40% noise rate in all types of inherent noisy labels. The best values of η for vanilla ODNL are reported in Table 5. C.1 Our method also works on other architectures.