- Keywords: Robustness, Augmentation, Geometric Prior
- TL;DR: We present elastically-augmented convolutions to improve the performance of CNNs, especially against natural perturbations
- Abstract: We focus on enhancing the performance of CNNs, especially against natural perturbations. Existing CNNs show outstanding performance on image classification, but training CNNs require large training datasets. We tackle this problem by incorporating elastic perturbations which approximate (local) view-point changes of the object in CNNs. We present elastically-augmented convolutions (EAConv) by parameterizing filters as a combination of fixed elastically-perturbed bases functions and trainable weights. We show an improvement in the general robustness of CNNs on the CIFAR-10 and STL-10 datasets, while even improving the performance on clean images without any data augmentation.