Research map / Label-efficient and robust learning / Test-time adaptation
Unified Entropy Optimization for Open-Set 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 UniEnt (and UniEnt+), a plug-in framework for open-set test-time adaptation that splits test batches into pseudo csID and csOOD samples via a feature-prototype cosine score with GMM, minimizing entropy on csID and maximizing it on csOOD.
Problem
Test-time adaptation methods tuned for closed-set covariate shift degrade both known-class accuracy and unknown-class detection when the unlabeled test stream contains samples from unseen classes under covariate shift.
Approach
A distribution-aware filter (cosine similarity between source-model features and class prototypes, modeled with a two-component GMM) separates csID from csOOD samples; entropy is minimized on csID and maximized on csOOD, with marginal entropy regularization. UniEnt+ replaces the hard split with sample-level weights (pi(x) and 1-pi(x)) to reduce partition noise.
Key results
- On CIFAR-10-C, UniEnt improves TENT's AUROC by 26.89 points and reduces FPR@TPR95 by 59.20 points, with accuracy gain of 4.93%.
- UniEnt+ on EATA over CIFAR-10-C raises AUROC from 84.32 to 96.97 and OSCR from 72.63 to 83.67.
- On Tiny-ImageNet-C, UniEnt improves TENT's OSCR from 19.02 to 30.18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add csID entropy minimization to TENT (ablation) | TENT · CIFAR-10-C | accuracy: 79.38 -> 85.04 | Helped |
| Add csOOD entropy maximization to TENT with csID entropy minimization (ablation) | TENT · CIFAR-10-C | AUROC: 81.80 -> 92.28 | Helped |
| Add csID entropy minimization to EATA (ablation) | EATA · CIFAR-10-C | accuracy: 80.92 -> 85.53 | Helped |
| Add csOOD entropy maximization to EATA with csID entropy minimization (ablation) | EATA · CIFAR-10-C | AUROC: 82.94 -> 97.15 | Helped |
| Add csID entropy minimization to OSTTA (ablation) | OSTTA · CIFAR-100-C | accuracy: 60.03 -> 58.95 | Hurt |
| Add csOOD entropy maximization to OSTTA with csID entropy minimization (ablation) | OSTTA · CIFAR-100-C | AUROC: 90.62 -> 94.84 | Helped |
| Apply UniEnt to EATA | EATA · CIFAR-100-C | accuracy: 60.63 -> 59.75 | Hurt |
| Apply UniEnt+ to EATA | EATA · CIFAR-100-C | AUROC: 88.64 -> 94.23 | Helped |
| Apply UniEnt to OSTTA | OSTTA · Tiny-ImageNet-C | accuracy: 37.29 -> 33.72 | Hurt |
| Apply UniEnt to OSTTA | OSTTA · Tiny-ImageNet-C | AUROC: 55.66 -> 62.69 | Helped |
| Vary λ1 from 0.1 to 1.0 (ablation) | TENT + UniEnt · CIFAR-100-C | accuracy: 59.09 -> 58.76 | Hurt |
| Vary λ2 from 0.1 to 1.0 (ablation) | TENT + UniEnt · CIFAR-100-C | AUROC: 87.02 -> 94.53 | Helped |
Surprising findings
- Entropy-based TTA methods such as TENT and OSTTA can perform worse than the no-adaptation Source and BN Adapt baselines under open-set conditions, with OSCR drops of 3.27% to 15.40% in some cases.
- Filtering csOOD samples with only the csID term (L_t,csID) also improves unknown-class detection, not just known-class accuracy.
Limitations
- Relies on a GMM split of a bimodal csOOD score, so hard partition noise remains in UniEnt, motivating UniEnt+.
- Evaluated only on corruption-based benchmarks with synthetic csOOD data (SVHN-C, ImageNet-O-C).
Experimental setup
Datasets: CIFAR-10-C CIFAR-100-C Tiny-ImageNet-C SVHN-C ImageNet-O-C Tiny-ImageNet
Benchmarks: CIFAR-10-C CIFAR-100-C Tiny-ImageNet-C
Metrics: Accuracy AUROC FPR@TPR95 OSCR
Baselines: Source BN Adapt CoTTA TENT EATA OSTTA
Builds on: TENT EATA OSTTA CoTTA T3A
Closest papers by meaning
- Test-Time Adaptation on Noisy Data via Model-Pruning-Based Filtering and Flatness-Aware Entropy Minimization (AAAI 2025 · similarity 0.90)
- COME: Test-time Adaption by Conservatively Minimizing Entropy (ICLR 2025 · similarity 0.89)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.89)
- TEA: Test-time Energy Adaptation (CVPR 2024 · similarity 0.88)
- Evaluation of Test-Time Adaptation Under Computational Time Constraints (ICML 2024 · similarity 0.88)
- Bayesian Weight Enhancement with Steady-State Adaptation for Test-time Adaptation in Dynamic Environments (ICML 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.