Research map / Label-efficient and robust learning / Test-time adaptation
L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer
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 L-TTA, a test-time adaptation method that updates only a reconstructed stem layer (DWT-based domain embedding plus Gaussian channel attention) by minimizing uncertainty instead of entropy, achieving low memory cost on CIFAR-10-C, ImageNet-C and Cityscapes-C.
Problem
Entropy-minimization test-time adaptation requires forward/backward passes through the whole network, incurring high memory and training cost, and it struggles to use single or small-batch data effectively.
Approach
Replace the entropy objective with a channel-wise uncertainty objective (NLL on a Gaussian channel attention) and update only the first (stem) layer, which is reconstructed with a 2D Haar DWT before the convolution and an IDWT after it so the architecture stays drop-in compatible.
Key results
- On CIFAR-10-C with ResNet-26 the method reaches 17.1% average error versus 22.5% for REALM, a 5.4% improvement over SOTA.
- On ImageNet-C with ResNet-50 the method reaches 64.9% error, slightly worse than the best baseline but using about 1.7% of SOTA memory and 5.9% of REALM's memory.
- On Cityscapes-C the method reaches 59.8 mIoU versus 55.2 for EcoTTA, using 0.4 MB versus 918 MB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GCAL to the stem layer (ablation) | L-TTA · CIFAR-10-C | prediction error (%): 30.8 -> 17.1 | Helped |
| add DEL to the stem layer (ablation) | L-TTA · CIFAR-10-C | prediction error (%): 17.8 -> 17.1 | Helped |
| add GCAL to the stem layer (ablation) | L-TTA · ImageNet-C | prediction error (%): 72.6 -> 64.9 | Helped |
| add DEL to the stem layer (ablation) | L-TTA · ImageNet-C | prediction error (%): 67.7 -> 64.9 | Helped |
| add GCAL to the stem layer (ablation) | L-TTA · Cityscapes-C | mIoU (%): 55.4 -> 59.8 | Helped |
| add DEL to the stem layer (ablation) | L-TTA · Cityscapes-C | mIoU (%): 58.1 -> 59.8 | Helped |
| use DWT level 2 instead of DWT level 1 (ablation) | L-TTA · CIFAR-10-C | prediction error (%): 20.5 -> 18.0 | Helped |
| use DWT level 2 instead of DWT level 1 (ablation) | L-TTA · CIFAR-100-C | prediction error (%): 39.3 -> 37.5 | Helped |
| use DWT level 2 instead of DWT level 1 (ablation) | L-TTA · ImageNet-C | prediction error (%): 65.2 -> 64.9 | Helped |
| change the SE reduction size (ablation) | L-TTA · CIFAR-10-C | prediction error (%): 18.0; 17.9; 19.2; 18.0 | Mixed |
| change the SE reduction size (ablation) | L-TTA · CIFAR-100-C | prediction error (%): 37.5; 39.2; 38.2; 39.4 | Mixed |
| change the SE reduction size (ablation) | L-TTA · ImageNet-C | prediction error (%): 67.4; 67.2; 66.5; 64.9 | Helped |
Surprising findings
- Minimizing uncertainty only in the stem layer outperforms entropy-based methods that update the full network on CIFAR-10-C, despite updating far fewer parameters.
- The method's advantage shrinks on datasets with many classes, with the authors attributing this to a curse of dimensionality for the stem-only update.
Limitations
- Performance gains diminish as the number of classes grows (e.g., ImageNet-C), attributed to limited capacity of the stem layer.
- Continuously minimizing uncertainty causes divergence on ImageNet-C, requiring interval re-initialization of the stem layer.
- Evaluated mainly on corruption benchmarks with ResNet backbones; Cityscapes-C is self-generated and not released.
Experimental setup
Datasets: ImageNet CIFAR-10-C CIFAR-100-C ImageNet-C Cityscapes-C ImageNet-Sketch ImageNet-R ImageNet-A
Benchmarks: CIFAR-10-C CIFAR-100-C ImageNet-C Cityscapes-C ImageNet-Sketch ImageNet-R ImageNet-A
Metrics: prediction error (%) mIoU (%) memory usage (MB) training latency
Baselines: BN STAT TENT MEMO SFT EATA SAR REALM EcoTTA DDA CoTTA
Builds on: TENT EcoTTA Squeeze-and-Excitation Gaussian YOLOv3
Closest papers by meaning
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.90)
- Test-time Correlation Alignment (ICML 2025 · similarity 0.90)
- Test-Time Adaptation with Binary Feedback (ICML 2025 · similarity 0.89)
- Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment (CVPR 2025 · similarity 0.88)
- Unified Entropy Optimization for Open-Set Test-Time Adaptation (CVPR 2024 · similarity 0.88)
- TEA: Test-time Energy Adaptation (CVPR 2024 · 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.