AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Surgical Fine-Tuning Improves Adaptation to Distribution Shifts

Yoonho Lee, Annie S Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, Chelsea Finn

ICLR 2023 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 46 citations (OpenAlex)

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Select layers using relative gradient norm instead of tuning all parametersFine-tuning pretrained models · CIFAR-Caccuracy: 79.9 -> 81.4Improved
Select layers using relative gradient norm instead of tuning all parametersFine-tuning pretrained models · ImageNet-Caccuracy: 50.7 -> 51.2Improved
Select layers using relative gradient norm instead of tuning all parametersFine-tuning pretrained models · Living-17accuracy: 92.8 -> 93.5Improved
Select layers using relative gradient norm instead of tuning all parametersFine-tuning pretrained models · Entity-30accuracy: 79.3 -> 80.6Improved
Select layers using relative gradient norm instead of tuning all parametersFine-tuning pretrained models · CIFAR-Flipaccuracy: 85.9 -> 87.7Improved
Select layers using relative gradient norm instead of tuning all parametersFine-tuning pretrained models · Waterbirdsaccuracy: 88.0 -> 88.0No clear effect
Select layers using relative gradient norm instead of tuning all parametersFine-tuning pretrained models · CelebAaccuracy: 82.2 -> 82.2No clear effect
Use Auto-SNR layer selection instead of tuning all parametersFine-tuning pretrained models · CIFAR-Flipaccuracy: 85.9 -> 17.3Hurt
Use Auto-RGN layer selection rather than manual cross-validation over blocksSurgical fine-tuning · CIFAR-Caccuracy: 82.8 -> 81.4Hurt
Tune early layers rather than all parameters for input-level shiftsSurgical fine-tuning · CIFAR-C and ImageNet-CaccuracyImproved
Tune middle layers rather than all parameters for feature-level shiftsSurgical fine-tuning · Living-17 and Entity-30accuracyImproved
Tune the last layer rather than all parameters for output-level shiftsSurgical fine-tuning · CIFAR-Flip, Waterbirds, and CelebAaccuracyImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy