CIFAR-10でチューニング 事前学習のデータ(ImageNet)を削る ImageNet を 7-8割程度まで削っても ・性能はあまり変化しない ・データの選択法によっては性能が上がる en larger datasets, scaling could improve further (e.g., dashed lines in A). E pruning (f = 1) are labeled with their best-fit power law scaling ⇠ ↵ ⌫. (N an asymptotic constant error E(P ! 1) = 1.1% is subtracted from each of he power law scaling more clearly.) B Frac. of ImageNet used for pretraining ResNet50 netuned on CIFAR-10 Test accuracy (%) 0.75 1.0 0.5 0.25 0 Figure 4: Data pruning improves trans learning. A: CIFAR-10 performance a ViT pre-trained on all of ImageNet2 and fine-tuned on different pruned subs of CIFAR-10 under the EL2N metric. CIFAR-10 performance of ResNet50s trained on different pruned subsets of I geNet1K and fine-tuned on all of CIFA 10.