Research map / Label-efficient and robust learning / Test-time adaptation
Test-time Correlation Alignment
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 Test-time Correlation Alignment (TCA), which builds a pseudo-source correlation from high-certainty test instances and aligns test feature correlations to it via a closed-form linear transform (LinearTCA), also usable as a plug-in boost (LinearTCA+) for existing TTA methods.
Problem
Test-time adaptation under distribution shift without source data cannot compute the source feature correlation needed for CORAL, and existing TTA methods overlook feature correlations, incur heavy backpropagation cost, and suffer domain forgetting.
Approach
Approximate the source correlation with the correlation of the k most certain (lowest prediction-uncertainty) test embeddings, theoretically bound the resulting error, and align test features to it with a linear transformation and mean shift, requiring no backpropagation or parameter updates.
Key results
- LinearTCA+ reaches 76.74% average accuracy on CIFAR-10-C versus 73.88% for the best baseline (TSD), and 69.58% average over the domain generalization benchmarks.
- LinearTCA uses about 4% of the peak GPU memory and 0.6% of the runtime of the best TTA baseline (AdaNPC) on ViT-B/16 CIFAR-10-C.
- On CLIP (PACS/OfficeHome/VLCS), LinearTCA achieves 97.36, 83.77 and 84.18 average accuracy, above WATT-S (96.08, 81.71, 81.59 average accuracy).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LinearTCA+ correlation alignment to the selected TTA baseline | TTA methods · PACS, OfficeHome, DomainNet, CIFAR-10-C, CIFAR-100-C, ImageNet-C | accuracy: 89.58 vs. 87.08; 70.87 vs. 68.87; 48.29 vs. 47.65; 76.74 vs. 73.88; 55.65 vs. 54.00; 50.71 vs. 50.58 | Helped |
| add LinearTCA correlation alignment without model updates | source model · PACS and OfficeHome | accuracy: 86.33 vs. 84.54; 70.12 vs. 68.37 | Helped |
| add LinearTCA correlation alignment without model updates | source model · CIFAR-10-C, CIFAR-100-C, ImageNet-C | accuracy: 8.48%, 3.67% and 4.51% over the source model | Mixed |
| add LinearTCA correlation alignment | TTA method on CLIP · PACS, OfficeHome, VLCS | accuracy (AVG): 1.28%, 2.08%, and 2.85% | Helped |
| add a three-layer MLP transformation instead of the linear transformation (ablation) | LinearTCA and LinearTCA+ · PACS and OfficeHome | accuracy (AVG): 81.84 -> 84.70; 88.77 -> 88.98; 62.01 -> 63.84; 64.27 -> 64.56 | Helped |
| add a two-layer MLP transformation instead of the linear transformation (ablation) | LinearTCA and LinearTCA+ · PACS and OfficeHome | accuracy (AVG): 81.84 -> 84.25; 88.77 -> 88.93; 62.01 -> 63.87; 64.27 -> 64.47 | Helped |
| use real source distributions for alignment (ablation) | LinearTCA and LinearTCA+ · PACS and OfficeHome | accuracy (AVG): 83.59 -> 84.15; 88.77 -> 89.15; 63.42 -> 64.35; 63.61 -> 64.64 | Helped |
| remove the linear transformation W (ablation) | LinearTCA · PACS, OfficeHome, CIFAR-10-C, CIFAR-100-C | source-domain accuracy after adaptation: 99.42 (+0.08) -> 99.35 (0.00); 93.87 (-0.53) -> 94.40 (0.00); 91.16 (-1.20) -> 92.36 (0.00); 67.35 (-3.04) -> 70.39 (0.0 | Mixed |
| set the pseudo-source size to k = 10 and test-domain sample size to k2 = 2400 (ablation) | LinearTCA · OfficeHome | accuracy | Helped |
| use batch size 1 (ablation) | LinearTCA · OfficeHome Art | accuracy: 2.16% | Helped |
| use batch size 1 instead of larger batch sizes (ablation) | LinearTCA+ · OfficeHome Art | accuracy: 0.824 vs. 59.96 | Hurt |
Surprising findings
- LinearTCA shows positive backward transfer on PACS, improving source-domain accuracy after adaptation (+0.08%).
- LinearTCA underperforms several advanced TTA baselines on CIFAR-10/100-C and ImageNet-C despite gains over the source model.
Limitations
- Linear transformation may not capture nonlinear distribution shifts; MLP variants give modest gains.
- Performance ceiling depends on backbone feature quality and how far high-certainty embeddings are from the true source.
- Only one main hyperparameter k, but results on some corruption benchmarks fall behind stronger baselines.
Experimental setup
Datasets: PACS OfficeHome DomainNet CIFAR-10-C CIFAR-100-C ImageNet-C VLCS
Benchmarks: PACS OfficeHome DomainNet CIFAR-10-C CIFAR-100-C ImageNet-C VLCS
Metrics: accuracy peak GPU memory running time accuracy after returning to source domain
Baselines: BN T3A AdaNPC TENT PLC EATA SAR TSD TIPI TEA WATT CLIPArTT
Builds on: CORAL TENT T3A EATA TEA
Closest papers by meaning
- L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer (NeurIPS 2024 · similarity 0.90)
- Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment (CVPR 2025 · similarity 0.89)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.89)
- Test-time Adaptation for Regression by Subspace Alignment (ICLR 2025 · similarity 0.89)
- Domain-Specific Block Selection and Paired-View Pseudo-Labeling for Online Test-Time Adaptation (CVPR 2024 · similarity 0.88)
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.88)
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.