Generalization to translation shifts: a study in architectures and augmentations

We study how effective data augmentation is at capturing the inductive bias of carefully designed network architectures for spatial translation invariance. We evaluate various image classification architectures (antialiased, convolutional, vision transformer, and fully connected MLP networks) and data augmentation techniques towards generalization to large translation shifts. We observe that: (a) without data augmentation, all architectures, including convolutional networks with antialiased modification suffer some degradation in performance when evaluated on translated test distributions. Understandably, both the in-distribution accuracy and degradation to shifts is significantly worse for non-convolutional models. (b) The robustness of performance is improved by even a minimal augmentation of $4$ pixel random crop across all architectures. In some instances, even $1-2$ pixel random crop is sufficient. This suggests that there is a form of meta generalization from augmentation. For non-convolutional architectures, while the absolute accuracy is still low with this basic augmentation, we see substantial improvements in robustness to translation shifts. (c) With a sufficiently advanced augmentation pipeline ($4$ pixel crop+RandAugmentation+Erasing+MixUp), all architectures can be trained to have competitive performance in terms of in-distribution accuracy as well as generalization to large translation shifts.