Research map / Label-efficient and robust learning / Test-time adaptation
NC-TTT: A Noise Constrastive Approach for Test-Time Training
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 NC-TTT, a test-time training method that adds a noise-contrastive auxiliary task (discriminating noisy in-distribution vs. OOD projected feature maps) and adapts a ResNet50 encoder on unlabeled target data.
Problem
Deep vision models degrade under domain shift at test time, and test-time training needs an unsupervised auxiliary objective that effectively captures source-domain information.
Approach
Uses noise-contrastive estimation to learn the source feature distribution: a discriminator is trained to separate in-distribution noisy feature maps (small sigma_s) from OOD noisy ones (larger sigma_o), and at test time the encoder is updated to make target features look in-distribution to this discriminator.
Key results
- On CIFAR-10-C (level 5), NC-TTT averages 82.43% accuracy vs. 53.82% for ResNet50 and 82.08% for ClusT3.
- On CIFAR-100-C (level 5), NC-TTT averages 57.68% vs. 56.70% for ClusT3 and 31.37% for ResNet50.
- On VisDA-C, NC-TTT reaches 62.71% vs. 61.91% for ClusT3 and 46.31% for ResNet50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NC-TTT instead of ClusT3 | ResNet50 · CIFAR-10-C Level 5 (average) | accuracy: 82.08 -> 82.43 | Helped |
| Use NC-TTT instead of ClusT3 | ResNet50 · CIFAR-100-C Level 5 (average) | accuracy: 56.70 -> 57.68 | Helped |
| Use NC-TTT instead of ClusT3 | ResNet50 · VisDA-C | accuracy: 61.91 -> 62.71 | Helped |
Surprising findings
- On CIFAR-10-C, NC-TTT is not the top method on several corruptions, e.g. Gaussian noise where it trails ClusT3 and TTT++, which the authors attribute to the auxiliary task not biasing the model toward any shift type.
Limitations
- Gains are modest on CIFAR-10-C for some corruptions and on VisDA-C relative to ClusT3 (about 0.8 points).
- Hyperparameters (sigma_s, sigma_o) are chosen per dataset via experiments and a derived beta rule, with the in-domain variance of 0 used for CIFAR-10/100-C.
- The mechanism by which auxiliary tasks learn domain-related information is not well understood, as noted by the authors.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CIFAR-10-C CIFAR-100-C VisDA-C
Benchmarks: CIFAR-10-C CIFAR-100-C VisDA-C
Metrics: accuracy (%)
Baselines: PTBN TENT LAME TTT TTT++ ClusT3
Builds on: Test-Time Training (TTT) Noise-Contrastive Estimation ClusT3 TTT++ TENT
Closest papers by meaning
- Noisy Test-Time Adaptation in Vision-Language Models (ICLR 2025 · similarity 0.89)
- PALM: Pushing Adaptive Learning Rate Mechanisms for Continual Test-Time Adaptation (AAAI 2025 · similarity 0.86)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.86)
- Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · similarity 0.86)
- BoostAdapter: Improving Vision-Language Test-Time Adaptation via Regional Bootstrapping (NeurIPS 2024 · similarity 0.85)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.85)
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.