We've got proven that by investing off two-D awareness for scale[^reference-sixty] and by picking out predictive attributes from the middle on the network, a sequence transformer might be aggressive with top rated convolutional nets for unsupervised image classification. You'll be co