Research map / Continual learning / Domain and prompt-based
Maintaining Consistent Inter-Class Topology in Continual Test-Time 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 Topological Consistency Adaptation (TCA), a continual test-time adaptation method that preserves inter-class centroid topology via uniformity, intra-class compactness and batch-imbalance weighting losses on top of a mean-teacher self-training framework.
Problem
Continual test-time adaptation suffers from error accumulation under continuously shifting domains because self-training distorts the inter-class feature topology learned on the source domain and intra-class feature uniformity degrades.
Approach
Builds an inter-class topological graph from augmented class centroids (EMA-updated) and applies a Gaussian-potential uniformity loss, an intra-class compactness loss, and a batch-imbalance weighting of centroid edges, added to symmetric cross-entropy self-training and an alignment loss.
Key results
- TCA reduces average error on CIFAR10-C to 14.7% (vs. 16.2% for CoTTA) and CIFAR100-C to 29.7% and ImageNet-C to 59.3%.
- On the CCC long-sequence benchmark TCA reaches 32.9% average accuracy, versus 32.6% for RDumb and 7.9% for CoTTA.
- Ablation shows the full combination of BTW, Lintra and Linter lowers CIFAR10-C error from 16.05% to 14.70%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TCA instead of DSS | CTTA on CIFAR10-C · CIFAR10-C | error rate (%): 16.0 -> 14.7 | Helped |
| Use TCA instead of RMT | CTTA on CIFAR100-C · CIFAR100-C | error rate (%): 30.4 -> 29.7 | Helped |
| Use TCA instead of RMT | CTTA on ImageNet-C · ImageNet-C | error rate (%): 60.2 -> 59.3 | Helped |
| Use TCA instead of RDumb | CTTA on CCC-benchmark · CCC-Easy | accuracy (%): 49.3±0.88 -> 49.1±0.35 | Hurt |
| Use TCA instead of RDumb | CTTA on CCC-benchmark · CCC-Medium | accuracy (%): 38.9±1.40 -> 39.5±0.53 | Helped |
| Use TCA instead of RDumb | CTTA on CCC-benchmark · CCC-Hard | accuracy (%): 9.6±1.60 -> 10.1±0.22 | Helped |
| Add Lalign only (ablation) | TCA on CIFAR10-C · CIFAR10-C | error rate (%): 16.05 -> 15.92 | Helped |
| Add Linter only (ablation) | TCA on CIFAR10-C · CIFAR10-C | error rate (%): 16.05 -> 15.57 | Helped |
| Add Lintra only (ablation) | TCA on CIFAR10-C · CIFAR10-C | error rate (%): 16.05 -> 15.68 | Helped |
| Add Lintra and Linter (ablation) | TCA on CIFAR10-C · CIFAR10-C | error rate (%): 16.05 -> 15.41 | Helped |
| Add BTW to Lintra and Linter (ablation) | TCA on CIFAR10-C · CIFAR10-C | error rate (%): 15.41 -> 14.85 | Helped |
| Add Lalign to BTW, Lintra, and Linter (ablation) | TCA on CIFAR10-C · CIFAR10-C | error rate (%): 14.85 -> 14.70 | Helped |
Surprising findings
- Using only the alignment loss gives minimal improvement, which the authors attribute to it acting as redundant augmentation.
- Ada: the TENT baseline reaches 90.4% error on CIFAR100-C, showing severe error accumulation over the long run.
Limitations
- Gains over the strongest baselines are modest (e.g., 14.7% vs 16.2% CoTTA on CIFAR10-C, and 32.9 vs 32.6 on CCC against RDumb).
- Relies on hyperparameters (lambda1=0.025, lambda2=0.15, alpha, beta) whose sensitivity is not extensively studied.
Experimental setup
Datasets: CIFAR10-C CIFAR100-C ImageNet-C CCC
Benchmarks: CIFAR10-C CIFAR100-C ImageNet-C CCC-benchmark
Metrics: classification error rate (%) classification accuracy (%)
Baselines: TENT AdaContrast CoTTA RMT DSS BeCoTTA ETA EATA SANTA RDumb
Builds on: CoTTA RMT Alignment and uniformity (Wang & Isola) AdaContrast
Closest papers by meaning
- Learning to Generate Gradients for Test-Time Adaptation via Test-Time Training Layers (AAAI 2025 · similarity 0.85)
- Effortless Active Labeling for Long-Term Test-Time Adaptation (CVPR 2025 · similarity 0.84)
- DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.83)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.83)
- Ranked Entropy Minimization for Continual Test-Time Adaptation (ICML 2025 · similarity 0.83)
- ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation (ICLR 2024 · 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.