Research map / Label-efficient and robust learning / Test-time adaptation
Test-Time Adaptation with Binary Feedback
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.
Introduces test-time adaptation with sparse binary (correct/incorrect) feedback and proposes BiTTA, an RL-style dual-path method combining feedback-guided adaptation on uncertain samples with MC-dropout agreement-based self-adaptation on confident samples.
Problem
Test-time adaptation fails under severe domain shifts when relying only on unlabeled data, while active TTA requiring full-class labels is expensive for annotators.
Approach
Replace full-class labels with binary correct/incorrect feedback on a few uncertain predictions, treat it as a non-differentiable RL reward (REINFORCE with MC-dropout policy), and combine it with agreement-based pseudo-labeling on confident unlabeled samples in one objective.
Key results
- BiTTA reaches 87.20% average accuracy on CIFAR10-C (severity 5) versus 81.09% for the best binary-feedback baseline SimATTA*, averaging about 13.3 points gain over baselines per the paper.
- On CIFAR100-C BiTTA achieves 62.49% average and on Tiny-ImageNet-C 40.85%, both highest among compared methods.
- MC-dropout calibration gives ECE 0.062 versus 0.100 for softmax confidence on CIFAR10-C.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine BFA and ABA rather than using BFA alone (ablation) | BiTTA on CIFAR10-C · CIFAR10-C | accuracy: 58.90% -> 87.20% | Helped |
| Combine BFA and ABA rather than using ABA alone (ablation) | BiTTA on CIFAR10-C · CIFAR10-C | accuracy: 82.64% -> 87.20% | Helped |
| Remove MC-dropout (ablation) | BiTTA · CIFAR10-C | accuracy: 2.56% accuracy drop | Hurt |
| Increase MC-dropout inferences from 1 to 4 (ablation) | BiTTA · CIFAR10-C | ECE: 0.142 -> 0.064 | Helped |
| Replace MC-dropout uncertainty estimation with augmentation (ablation) | BiTTA · CIFAR10-C | accuracy: 87.20 -> 19.07 | Hurt |
| Replace MC-dropout uncertainty estimation with an ensemble (ablation) | BiTTA · CIFAR10-C | accuracy: 87.20 -> 85.50 | Hurt |
| Replace MC-dropout uncertainty estimation with deterministic confidence (ablation) | BiTTA · CIFAR10-C | accuracy: 87.20 -> 85.90 | Hurt |
| Use 1 adaptation epoch instead of 3 (ablation) | BiTTA · CIFAR10-C | accuracy: 87.20 -> 87.11 | No clear effect |
| Use 2 adaptation epochs instead of 3 (ablation) | BiTTA · CIFAR10-C | accuracy: 87.20 -> 87.15 | No clear effect |
| Apply MECTA memory-efficient continual TTA (ablation) | BiTTA · CIFAR10-C | GPU memory: 8304 -> 6724 | Helped |
| Use gradient accumulation (ablation) | BiTTA · CIFAR10-C | GPU memory: 8304 -> 2841 | Helped |
| Use BiTTA instead of SAR* | Binary-feedback TTA on CIFAR10-C · CIFAR10-C severity 5 | accuracy: 83.78 -> 87.20 | Helped |
Surprising findings
- Under equal labeling cost, binary-feedback BiTTA outperforms full-class SimATTA labels by up to 32% accuracy.
- Replacing oracle labels with GPT-4o as feedback substantially degrades SimATTA because GPT-4o has about 40% error under shift.
Limitations
- Still requires substantial oracle involvement in high-variability or rapid shift settings.
- Higher computational and memory overhead from repeated MC-dropout forward passes.
- Does not explicitly handle noisy or incorrect binary feedback; accuracy degrades under feedback errors.
Experimental setup
Datasets: CIFAR10-C CIFAR100-C Tiny-ImageNet-C PACS ImageNet-C ImageNet-R ColoredMNIST VisDA-2021 DomainNet
Benchmarks: CIFAR10-C CIFAR100-C Tiny-ImageNet-C PACS ImageNet-C ImageNet-R ColoredMNIST VisDA-2021 DomainNet
Metrics: accuracy expected calibration error wall-clock time per batch GPU memory
Baselines: BN-Stats TENT EATA SAR CoTTA RoTTA SoTTA SimATTA DeYO TAST-BN OWTTT
Builds on: TENT SimATTA REINFORCE MC-dropout
Closest papers by meaning
- L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer (NeurIPS 2024 · similarity 0.89)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.89)
- Improved Self-Training for Test-Time Adaptation (CVPR 2024 · similarity 0.88)
- Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization (AAAI 2024 · similarity 0.87)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.87)
- AETTA: Label-Free Accuracy Estimation for Test-Time Adaptation (CVPR 2024 · similarity 0.87)
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.