Training process is crucial for the deployment of the network in applications which have two strict requirements on both accuracy and robustness. However, most existing approaches are in a dilemma, i.e. model accuracy and robustness forming an embarrassing tradeoff – the improvement of one leads to the drop of the other. The challenge remains for as we try to improve the accuracy and robustness simultaneously. In this paper, we propose a novel training method via introducing the auxiliary classifiers for training on corrupted samples, while the clean samples are normally trained with the primary classifier. In the training stage, a novel distillation method named input-aware self distillation is proposed to facilitate the primary classifier to learn the robust information from auxiliary classifiers. Along with it, a new normalization method - selective batch normal- ization is proposed to prevent the model from the negative influence of corrupted images. At the end of the training period, a L2-norm penalty is applied to the weights of primary and auxiliary classifiers such that their weights are asymptotically identical. In the stage of inference, only the primary classifier is used and thus no extra computation and storage are needed. Extensive experiments on CIFAR10, CIFAR100 and ImageNet show that noticeable improvements on both accuracy and robustness can be observed by the proposed auxiliary training. On average, auxiliary training achieves 2.21% accuracy and 21.64% robustness (measured by corruption error) improvements over traditional training methods on CIFAR100. Codes have been released on github.

No abstract uploaded!

We show that a properly immersed minimal hypersurface in M × R+ equals some M×{c} when M is a complete, recurrent ndimensional Riemannian manifold with bounded curvature. If on the other hand,M is not necessarily recurrent but has nonnegative Ricci curvature with curvature bounded below, the same result holds for any positive entire minimal graph over M.

We investigate the angular distribution of Ly$\alpha$ photons transferring in or emergent from an optically thick medium. Since the evolutions of specific intensity $I$ in the frequency space and the angular space are coupled with each other, we first develop the WENO numerical solver in order to find the time-dependent solutions of the integro-differential equation of $I$ in frequency and angular space simultaneously. We first show that the solutions with the Eddington approximation, which assume $I$ to be linearly dependent on the angular variable $\mu$, yield similar frequency profiles of the photon flux as that without the Eddington approximation. However, the solutions of the $\mu$ distribution evolution are significantly different from that given by Eddington approximation. First, the angular distribution of $I$ are found to be substantially dependent on the frequency of photons. For photons with the resonant frequency $\nu_0$, $I$ contains only a linear term of $\mu$. For photons with frequency at the double peaks of the flux, the $\mu$-distribution is highly anisotropic, in which most photons are in the direction of radial forward. Moreover, either at $\nu_0$ or at the double peaks, the $\mu$ distributions actually are independent of the initial $\mu$ distribution of photons of the source. This is because the photons with frequency either of $\nu_0$ or of the double peaks undergo the process of forgetting their initial conditions due to the resonant scattering. We also show that the optically thick medium is a collimator of photons at the double peaks. Photons from the double peaks form a forward beam with very small spread angle.

No abstract uploaded!