Research map / Label-efficient and robust learning / Test-time adaptation
COME: Test-time Adaption by Conservatively Minimizing Entropy
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 COME, a drop-in replacement for entropy minimization in test-time adaptation that minimizes the entropy of a subjective-logic (Dirichlet) opinion with a detached-norm constraint, improving accuracy and FPR across standard, open-world and lifelong TTA on ImageNet-C.
Problem
Entropy minimization in test-time adaptation is overconfident and unstable, leading to model collapse when test samples are unreliable or outliers occur.
Approach
Model the classifier output as Dirichlet evidence via subjective logic so that an uncertainty mass exists; minimize the entropy of the resulting opinion (belief masses plus uncertainty) while detaching the logit norm so the uncertainty is regularized, giving a data-adaptive upper bound on confidence.
Key results
- With SAR on ImageNet-C level 5, COME raises average accuracy from 54.2% to 64.2% and reduces FPR from 66.7% to 63.8%.
- In lifelong TTA, COME improves SAR accuracy by up to 34.8% on average and reduces FPR by 15.1%.
- Applied to Tent in open-world TTA, COME improves average accuracy by 2.4% (68.6 to 71.0).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace entropy minimization with COME in SAR | SAR · ImageNet-C level 5 | accuracy: +10.1 | Helped |
| Replace entropy minimization with COME in SAR | SAR · ImageNet-C level 5 | FPR: +2.9 | Helped |
| Replace entropy minimization with COME in Tent | Tent · ImageNet-C level 5 | accuracy: +8.4 | Helped |
| Replace entropy minimization with COME in EATA | EATA · ImageNet-C level 5 | accuracy: +2.4 | Helped |
| Replace entropy minimization with COME in CoTTA | CoTTA · ImageNet-C level 5 | accuracy: +3.0 | Helped |
| Replace entropy minimization with COME in MEMO | MEMO · ImageNet-C level 5 | accuracy: +0.9 | Helped |
| Replace entropy minimization with COME in SAR | SAR · Open-world TTA | accuracy: +5.3 | Helped |
| Replace entropy minimization with COME in EATA | EATA · Open-world TTA | FPR: +8.6 | Helped |
| Replace entropy minimization with COME in CoTTA | CoTTA · Open-world TTA | FPR | No clear effect |
| Replace entropy minimization with COME in SAR | SAR · Lifelong ImageNet-C | accuracy: +34.8 | Helped |
| Replace entropy minimization with COME in SAR | SAR · Lifelong ImageNet-C | FPR: +15.1 | Helped |
| Replace entropy minimization with COME in MEMO | MEMO · ImageNet validation | accuracy | No clear effect |
Surprising findings
- Non-EM methods such as PL and TEA are generally less effective than EM-based methods in the reported ImageNet-C results, despite avoiding entropy minimization.
Limitations
- Theoretical analysis is described as preliminary; the stability result is an upper bound on confidence rather than a convergence guarantee.
- Hyperparameters p=2 and tau=1 are fixed for simplicity without a broader sensitivity study in the main text.
- Gains on some corruptions and metrics are mixed; some FPR values degrade relative to EM baselines (e.g., Tent/EATA FPR under certain open-world outlier sets).
Experimental setup
Datasets: ImageNet-C ImageNet-1K OpenImage-O SSB-Hard Textures iNaturalist NINCO Places
Benchmarks: ImageNet-C ImageNet-C open-world (NINCO, iNaturalist, SSB-Hard, Textures, Places)
Metrics: accuracy FPR95
Baselines: Tent EATA SAR CoTTA MEMO PL T3A TEA LAME FOA
Builds on: Tent EATA SAR CoTTA subjective logic
Closest papers by meaning
- Unified Entropy Optimization for Open-Set Test-Time Adaptation (CVPR 2024 · similarity 0.89)
- Protected Test-Time Adaptation via Online Entropy Matching: A Betting Approach (NeurIPS 2024 · similarity 0.87)
- TEA: Test-time Energy Adaptation (CVPR 2024 · similarity 0.87)
- Ranked Entropy Minimization for Continual Test-Time Adaptation (ICML 2025 · similarity 0.87)
- Entropy is not Enough for Test-Time Adaptation: From the Perspective of Disentangled Factors (ICLR 2024 · similarity 0.86)
- Test-Time Adaptation on Noisy Data via Model-Pruning-Based Filtering and Flatness-Aware Entropy Minimization (AAAI 2025 · similarity 0.86)
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.