Research map / Label-efficient and robust learning / Test-time adaptation
Beyond Entropy: Region Confidence Proxy for Wild 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 ReCAP, which replaces entropy minimization in wild test-time adaptation with a closed-form upper-bounded proxy of region confidence (regional entropy plus regional instability) over a Gaussian feature neighborhood.
Problem
Entropy minimization in Wild Test-Time Adaptation (limited batch size, mixed domains, imbalanced label shift) produces noisy, locally inconsistent optimization dynamics, and the region-level alternative is computationally intractable.
Approach
Model each test feature's local region as a Gaussian with source-estimated diagonal variance, then derive closed-form upper bounds (Regional Entropy and Regional Instability) via finite-to-infinite asymptotic approximation, so region confidence can be optimized without sampling or extra forward passes; the proxy is also used to select and weight reliable samples.
Key results
- Under limited batch size 1 on ImageNet-C (severity 5), ReCAP reaches 46.4% average accuracy with ResNet50 and 65.7% with ViT-Base, versus 44.1% and 64.1% for the previous best DeYO.
- ReCAP+DeYO reaches 48.9% (ResNet50) and 65.9% (ViT-Base) on the same setting, the best combined numbers in the table.
- On ImageNet-C under mixed testing domain, ReCAP achieves 46.4% average versus 44.4% for DeYO with ResNet50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace entropy minimization with ReCAP | ResNet50 · ImageNet-C severity 5, limited batch size 1 | accuracy: 44.1 -> 46.4 | Helped |
| Replace entropy minimization with ReCAP | ViTBase · ImageNet-C severity 5, limited batch size 1 | accuracy: 64.1 -> 65.7 | Helped |
| Replace entropy minimization with ReCAP | ResNet50 · ImageNet-C severity 5, imbalanced label shift | accuracy: +1.5% | Helped |
| Replace entropy minimization with ReCAP | ViTBase · ImageNet-C severity 5, imbalanced label shift | accuracy: +2.2% | Helped |
| Integrate ReCAP with SAR | SAR · Three wild scenarios | accuracy: +11.0%, +3.1%, and +5.9% | Helped |
| Integrate ReCAP with DeYO | DeYO · Three wild scenarios | accuracy: +3.3%, +1.6%, and +4.0% | Helped |
| Remove regional instability, retaining regional entropy only (ablation) | ReCAP · ImageNet-C imbalanced label shift, corruption categories average | accuracy: 45.8 -> 43.3 | Hurt |
| Remove regional entropy and use vanilla entropy for selection and weighting (ablation) | ReCAP · ImageNet-C imbalanced label shift, corruption categories average | accuracy: 45.8 -> 31.4 | Hurt |
| Replace regional entropy with vanilla entropy selection and weighting, retaining regional instability (ablation) | ReCAP · ImageNet-C imbalanced label shift, corruption categories average | accuracy: 45.8 -> 42.7 | Hurt |
| Increase the variance-term weight beyond the optimal range (ablation) | ReCAP · ImageNet-C imbalanced label shift | accuracy | Hurt |
| Increase the local-region scale beyond a reasonable range, e.g. τ=2.5 (ablation) | ReCAP · ImageNet-C imbalanced label shift | accuracy | Hurt |
Surprising findings
- Tent collapses under wild settings, with ResNet50 average accuracy of 21.5% under batch size 1, far below the 30.6% no-adapt baseline.
Limitations
- Evaluation is limited to image classification benchmarks; other tasks are not explored.
- The local region is domain-wise with a shared shape for all samples, not class-wise or sample-wise adaptive.
Experimental setup
Datasets: ImageNet-C ImageNet-R VisDA-2021
Benchmarks: ImageNet-C ImageNet-R VisDA-2021
Metrics: accuracy
Baselines: Tent EATA SAR DeYO MEMO DDA
Builds on: Tent SAR DeYO EATA
Closest papers by meaning
- Unified Entropy Optimization for Open-Set Test-Time Adaptation (CVPR 2024 · similarity 0.88)
- Test-Time Adaptation on Noisy Data via Model-Pruning-Based Filtering and Flatness-Aware Entropy Minimization (AAAI 2025 · similarity 0.88)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.87)
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.87)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.87)
- Towards Robust Multimodal Open-set Test-time Adaptation via Adaptive Entropy-aware Optimization (ICLR 2025 · 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.