Long-ShortTransformer: EfficientTransformers forLanguageandVision(Appendix) ADetailsofNormComparisons

The first design helps the model focus more on the global context of the image as each patch could attend to the whole image areas. It reduces the local texture bias ofCNN.