Research map / Label-efficient and robust learning / Test-time adaptation
Online Feature Updates Improve Online (Generalized) Label Shift Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OLS-OFU, which couples self-supervised feature-extractor updates (batch-accumulated every τ steps) with existing online label shift methods, improving error on CIFAR-10 and generalized label shift benchmarks while keeping regret guarantees.
Problem
Online label shift and online generalized label shift adaptation with missing labels: existing methods only reweight or update the final linear layer of a frozen pretrained classifier, leaving the feature extractor un-adapted at test time.
Approach
Insert a self-supervised feature-extractor update (on accumulated unlabeled test batches every τ steps) after the revised online label shift step, then retrain the last linear layer on training data, preserving unbiased gradient estimates and regret guarantees of the base OLS method.
Key results
- On CIFAR-10 with sinusoidal shift, OLS-OFU (tau=100) lowers average error for all six OLS methods, e.g., FLH-FTL from 12.02% to 7.45% and FTFWH from 11.9% to 7.33%.
- Batch accumulation with tau=100 reduces the added time cost by about 99% compared with per-step feature updates (tau=1), e.g., FLH-FTL-OFU time 4.15 min vs 285 min for tau=1.
- Even with 0% stored training data for linear retraining, OLS-OFU (FTH) still achieves 10.24% vs 12.04% for FTH alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add online feature updates with rotation degree prediction at τ=100 | OLS methods · CIFAR-10 online label shift | average error: 11.9% -> 7.33%; 12.04% -> 8%; 13.65% -> 10.8%; 12.18% -> 10.1%; 11.54% -> 8.35%; 12.02% -> 7.45% | Helped |
| Update the feature extractor before OLS instead of after OLS (ablation) | OLS-OFU · CIFAR-10 online label shift, sinusoidal | average error: 11.3% -> 12.33%; 11.2% -> 12.12%; 13.9% -> 14.35%; 11.6% -> 12.10%; 11.4% -> 11.91%; 11.2% -> 12.08% | Hurt |
| Update the feature extractor before OLS instead of after OLS (ablation) | OLS-OFU (τ=100) · CIFAR-10 online label shift, sinusoidal | average error: 7.23% -> 8.02%; 7.91% -> 8.63%; 10.81% -> 11.19%; 10.12% -> 10.55%; 8.33% -> 8.80%; 7.53% -> 8.29% | Hurt |
| Increase batch-accumulation frequency from τ=1 to τ=100 (ablation) | OLS-OFU with rotation degree prediction · CIFAR-10 online label shift | average error: 11.3% -> 7.33%; 11.2% -> 8%; 13.9% -> 10.8%; 11.6% -> 10.1%; 11.4% -> 8.35%; 11.2% -> 7.45% | Helped |
| Increase batch-accumulation frequency from τ=100 to τ=500 (ablation) | OLS-OFU with rotation degree prediction · CIFAR-10 online label shift | average error: 7.33% -> 11.6%; 8% -> 11.7%; 10.8% -> 13.5%; 10.1% -> 12.1%; 8.35% -> 11.3%; 7.45% -> 11.7% | Hurt |
| Reduce stored training data for retraining the linear layer from 100% to 0% (ablation) | OLS-OFU using FTH, ROGD, FLHFTL, or FTFWH · CIFAR-10 online label shift, sinusoidal | average error: 8% -> 10.24%; 10.8% -> 13.50%; 7.45% -> 10.43%; 7.33% -> 10.41% | Hurt |
| Add feature updates while using 0% stored training data and reusing the pretrained linear classifier (ablation) | OLS-OFU using FTH, ROGD, FLHFTL, or FTFWH · CIFAR-10 online label shift, sinusoidal | average error: 10.24% vs 12.04%; 13.50% vs 13.65%; 10.43% vs 12.02%; 10.41% vs 11.9% | Helped |
| Increase the amount of stored training data used to retrain the last linear layer (ablation) | OLS-OFU using FTH, ROGD, FLHFTL, or FTFWH · CIFAR-10 online label shift, sinusoidal | average error | Helped |
Surprising findings
- Updating the feature extractor without retraining the linear layer (0% stored training data) still outperforms the original OLS methods.
Limitations
- Retraining the linear layer requires storing training data at test time, adding memory and compute cost.
- Regret guarantees rely on label-shift assumptions and known or estimated confusion matrices; generalized label shift is only argued heuristically via test-time training.
- Per-step feature updates are very expensive (247x time for FLH-FTL), requiring batch accumulation.
Experimental setup
Datasets: CIFAR-10 STL10 CINIC EuroSAT CIFAR-10C
Benchmarks: CIFAR-10 STL10 CINIC EuroSAT CIFAR-10C
Metrics: average error wall-clock time
Baselines: FTFWH FTH ROGD ATLAS UOGD FLH-FTL Base (fixed pre-trained model)
Builds on: test-time training FLH-FTL ROGD UOGD ATLAS
Closest papers by meaning
- Label Shift Meets Online Learning: Ensuring Consistent Adaptation with Universal Dynamic Regret (CVPR 2025 · similarity 0.88)
- Learning Label Shift Correction for Test-Agnostic Long-Tailed Recognition (ICML 2024 · similarity 0.84)
- Heterogeneous Label Shift: Theory and Algorithm (ICML 2025 · similarity 0.83)
- Protected Test-Time Adaptation via Online Entropy Matching: A Betting Approach (NeurIPS 2024 · similarity 0.83)
- Online Boosting Adaptive Learning under Concept Drift for Multistream Classification (AAAI 2024 · similarity 0.83)
- Rethinking Self-Distillation: Label Averaging and Enhanced Soft Label Refinement with Partial Labels (ICLR 2025 · similarity 0.82)
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.