Long-Tail Learning with Foundation Model: Heavy Fine-Tuning Hurts

Open paper
Benchmark Model Rank Results
long-tail-learning-on-cifar-100-lt-r-10LIFT (ViT-B/16, ImageNet-21K pre-training)#1Error Rate: 8.7
long-tail-learning-on-cifar-100-lt-r-10LIFT (ViT-B/16, CLIP)#4Error Rate: 15.1
long-tail-learning-on-cifar-100-lt-r-100LIFT (ViT-B/16, ImageNet-21K pre-training)#2Error Rate: 10.9
long-tail-learning-on-cifar-100-lt-r-100LIFT (ViT-B/16, CLIP)#3Error Rate: 18.3
long-tail-learning-on-cifar-100-lt-r-50LIFT (ViT-B/16, ImageNet-21K pre-training)#1Error Rate: 9.8
long-tail-learning-on-cifar-100-lt-r-50LIFT (ViT-B/16, CLIP)#4Error Rate: 16.9
long-tail-learning-on-imagenet-ltLIFT (ViT-L/14)#1Top-1 Accuracy: 82.9
long-tail-learning-on-imagenet-ltLIFT (ViT-B/16)#3Top-1 Accuracy: 78.3
long-tail-learning-on-inaturalist-2018LIFT (ViT-L/14@336px)#1Top-1 Accuracy: 87.4%
long-tail-learning-on-inaturalist-2018LIFT (ViT-L/14)#2Top-1 Accuracy: 85.2%
long-tail-learning-on-inaturalist-2018LIFT (ViT-B/16)#5Top-1 Accuracy: 80.4%
long-tail-learning-on-places-ltLIFT (ViT-L/14)#1Top-1 Accuracy: 53.7
long-tail-learning-on-places-ltLIFT (ViT-B/16)#2Top-1 Accuracy: 52.2

Rank counts only results with a code link.