Research map / Label-efficient and robust learning / Test-time adaptation
A Layer Selection Approach to 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 GALA, a per-sample layer selection criterion for test-time adaptation that updates only layers whose gradient updates are angularly aligned with the accumulated parameter displacement (with periodic anchor resets), improving PL and SHOT losses across DomainBed and Continual TTA benchmarks.
Problem
Test-time adaptation under distribution shift often degrades performance below the pretrained model, partly because adapting certain layers with misaligned gradients harms accuracy, and existing layer choices are fixed across shifts and losses.
Approach
Rank layers per sample by the cosine between the proposed update and the anticipated total displacement from the pretrained (anchor) parameters, updating only layers above a threshold; the anchor is reset every fixed window so the criterion can track direction changes.
Key results
- On DomainBed with ResNet-50 and PL loss, GALA reaches 69.80% mean accuracy vs 67.93% for ERM and 63.85% for all-layers TTA.
- On Continual TTA, GALA with SHOT reduces CIFAR10-C error to 20.46% vs 89.33% for all-layers SHOT.
- On Tiny-Domainbed, GALA achieves 59.4% TTA accuracy and 9.3% forgetting, with Spearman correlation 0.76 to the best-block oracle.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GALA layer selection instead of all-layer adaptation | ResNet-18 with PL · Domainbed Mean | accuracy: 60.77 -> 66.47 | Helped |
| Use GALA layer selection instead of ERM | ResNet-18 with PL · Domainbed Mean | accuracy: 64.78 -> 66.47 | Helped |
| Use GALA layer selection instead of all-layer adaptation | ResNet-18 with SHOT · Domainbed Mean | accuracy: 61.64 -> 66.40 | Helped |
| Use GALA layer selection instead of all-layer adaptation | ResNet-50 with PL · Domainbed Mean | accuracy: 63.85 -> 69.80 | Helped |
| Use GALA layer selection instead of all-layer adaptation | ResNet-50 with SHOT · Domainbed Mean | accuracy: 63.03 -> 69.17 | Helped |
| Use GALA layer selection instead of all-layer adaptation | Continual TTA with PL · CIFAR10C | error rate: 88.72~(\pm 1.2) -> 28.68~(\pm 6.6) | Helped |
| Use GALA layer selection instead of all-layer adaptation | Continual TTA with PL · CIFAR100C | error rate: 98.63~(\pm 1.5) -> 33.69~(\pm 5.7) | Helped |
| Use single-layer rather than single-block selection (ablation) | GALA · Domainbed averaged settings | accuracy: 67.64 -> 68.57 | Helped |
| Select multiple layers rather than one layer (ablation) | GALA · Domainbed averaged settings | accuracy: 68.57 -> 66.48 | Hurt |
| Change reset window size from 20 to 5 (ablation) | GALA · Domainbed averaged settings | accuracy: 68.57 -> 68.46 | No clear effect |
| Disable reset (ablation) | GALA · Continual TTA | accuracy: 69.9 -> 71.1 | Hurt |
| Set selection threshold to 0.99 instead of 0.75 (ablation) | GALA · Domainbed averaged settings | accuracy: 68.57 -> 68.6 | No clear effect |
Surprising findings
- Adapting all layers can degrade accuracy well below the frozen pretrained model (e.g., CIFAR10-C error near 89% for all-layers TTA vs 43.5% for ERM).
- The best adapted layer can change with the TTA loss even for the same shift and dataset.
Limitations
- Gains on some DomainBed settings are within large standard deviations (e.g., TerraIncognita).
- GALA is more conservative than the best-block oracle, sometimes stopping useful updates and reducing TTA accuracy.
- Experiments are limited to image classification with ResNet and WideResNet/ResNeXt backbones.
Experimental setup
Datasets: PACS VLCS TerraIncognita OfficeHome CIFAR10-C CIFAR100-C
Benchmarks: DomainBed Continual TTA (CIFAR10-C, CIFAR100-C)
Metrics: accuracy error rate forgetting generalization accuracy Spearman rank correlation
Baselines: ERM All layers AutoRGN AutoSNR
Builds on: pseudo-labeling SHOT AutoRGN AutoSNR T3A
Closest papers by meaning
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.92)
- Buffer layers for Test-Time Adaptation (NeurIPS 2025 · similarity 0.92)
- Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization (AAAI 2024 · similarity 0.91)
- Evaluation of Test-Time Adaptation Under Computational Time Constraints (ICML 2024 · similarity 0.90)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.90)
- L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer (NeurIPS 2024 · similarity 0.90)
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.