Notably, it achieves 75.7% top-1 accuracy with only 1% of ImageNet-1K
labels when fine-tuned.