IC-1017Vision and language models pre-trained on noisy data exhibit degraded OOD transfer that is partially recoverable via SVD-based feature-space regularization

Hao Chen, Jindong Wang, Ankit Shah, Ran Tao, Hongxin Wei, Xing Xie, Masashi Sugiyama, Bhiksha Raj

SourceUnderstanding and Mitigating the Label Noise in Pre-training on Downstream Tasks

The paper evaluates nine released models (EfficientNet-B3, ResNetV2-152x2, Swin-L, ViT-L, ConvNeXt-L, BERT-L, RoBERTa-L, GPT-2, text-ada-002) that are known to be pre-trained on noisy or uncurated data (JFT-300m, ImageNet-21k, LAION-2B, WebText, BookCorpus+Wikipedia). Under linear probing, all models show a substantial gap between in-domain and out-of-domain accuracy. Applying the authors' NMtune method (a 2-layer MLP with MSE, covariance, and dominant-singular-value regularization losses) improves OOD accuracy over LP by 1–9 points across all models, with the largest gains on language models (e.g., text-ada-002 GLUE-X: 44.06→53.48) and vision models (e.g., ConvNeXt-L DomainNet: 66.86→70.30). MLP tuning without the regularization losses provides smaller or inconsistent gains, indicating the improvement comes from reshaping the feature-space spectrum rather than from added capacity.

Evidence
correlational
Key metric
OOD accuracy (avg): Swin-L 50.88 (LP) → 52.35 (NMtune); ConvNeXt-L 66.86 → 70.30; GPT-2 GLUE-X 36.68 → 39.07; text-ada-002 GLUE-X 44.06 → 53.48; BERT-L GLUE-X 50.65 → 51.63; RoBERTa-L GLUE-X 44.55 → 47.01
Caveat
The paper notes that results are not comparable to white-box full fine-tuning, and that the explored models are relatively small-scale in terms of pre-training. The authors also note that NMtune may not be noise-tolerant on the downstream side (performance falls short of LP when downstream noise exceeds 40%).
Model
EfficientNet EfficientNet-B3, ResNet / ResNet-152 / ResNet-101 / ResNet-50-BN ResNetV2-152x2, Swin Transformer Swin-L, ViT ViT-L, ConvNeXt ConvNeXt-L, BERT BERT-L, RoBERTa / RoBERTa-L, GPT-2, text-ada-002
Datasets
GLUE [eval], GLUE-X [eval], DomainNet [eval], ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval], CIFAR-10 [eval], CIFAR-100 [eval]
Methods
Linear Probing / Ridge regression linear probing / Linear probe / Linear probe fine-tuning / Linear regression probing / Linear ridge regression probes / Supervised probing / ERM linear probe [compared-to]
Extraction
automatic-extraction