Research map / Label-efficient and robust learning / Recognition and long-tail
Stabilizing Feature Geometry in Noisy Pretrained Models for Robust Downstream Tasks
Quanyu Zhang, Zhongyi Han, Hao Sun, Yongshun Gong, Xiaoyan Wang, Yilong Yin, Shuo Li
Read the paperCheck your idea against 137,217 papers
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add Feature Geometry Stabilization on top of MLP tuning | MLP probe on noisy-pretrained backbones · DomainNet OOD, ViT-L Laion-2B | accuracy: 0.7359 -> 0.7662 | Improved |
| add FGS instead of NMTune spectral regularization | ResNetv2-152x2 ImageNet-21K · DomainNet OOD | accuracy: 0.5052 -> 0.5408 | Improved |
| NMTune spectral amplification versus plain MLP | ResNetv2-152x2 ImageNet-21K · DomainNet OOD | accuracy: 0.5126 (MLP) -> 0.5052 (NMTune) | Hurt |
| add FGS on in-domain tasks | ResNetv2-152x2 ImageNet-21K · 8 in-domain datasets | accuracy: 0.8523 -> 0.8665 (MLP), 0.8511 (NMTune) | Improved |
| add Feature Perturbation Consistency alone (ablation) | MLP probe · OOD DomainNet Sketch, ResNet-50 on noisy ImageNet-1K and YFCC15M | accuracy: FPC beats VAR by 1.29% and FCD by 0.88% at 30% noise on ImageNet-1K | Improved |
| add VAR alone or FCD alone (ablation) | MLP probe · OOD DomainNet Sketch, ResNet-50 | accuracy | Improved |
| combine FPC, VAR and FCD (ablation) | FGS · OOD DomainNet Sketch, ResNet-50 all noise levels | accuracy | Improved |
| use salt-and-pepper, Gaussian or Poisson feature perturbation in FPC (ablation) | FPC · OOD DomainNet, ResNet-50 | accuracy | Improved |
| use generator-based synthesized noise for FPC perturbation (ablation) | FPC · OOD DomainNet, ResNet-50 | accuracy | Hurt |
| apply FGS under dual noise (pretraining and downstream label noise) | ResNet-50 YFCC15M · CIFAR-100 with downstream noise 0-50% | accuracy: drop 12-23% smaller than NMTune at 30% pretraining noise | Improved |
| increase pretraining label noise ratio | ResNet-50 linear probe · 12 downstream tasks | accuracy vs principal directional angle | Hurt |
Closest papers by meaning
- Stabilizing Streaming Video Geometry via Dynamic Feature Normalization (CVPR 2026 · similarity 0.87)
- Which Models have Perceptually-Aligned Gradients? An Explanation via Off-Manifold Robustness (NeurIPS 2023 · similarity 0.87)
- Restoring Initial Noise Sensitivity in Text-to-Image Distillation through Geometric Alignment (ICML 2026 · similarity 0.85)
- Flow-Corrected Shape Optimization: Taming Manifold Drift in High-Dimensional 3D Models (NeurIPS 2026 · similarity 0.84)
- Understanding and Mitigating the Label Noise in Pre-training on Downstream Tasks (ICLR 2024 · similarity 0.84)
- Beyond Accuracy: Ensuring Correct Predictions With Correct Rationales (NeurIPS 2024 · similarity 0.84)
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.
