Research map / Label-efficient and robust learning / Test-time adaptation
Cross-Device Collaborative 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 CoLA, a cross-device collaborative lifelong test-time adaptation paradigm that stores device-shared domain knowledge vectors and uses either backprop-based knowledge reprogramming (principal agents) or forward-only similarity-based aggregation (follower agents) on ImageNet-C and related benchmarks.
Problem
Test-time adaptation is performed independently per device, ignoring knowledge from other devices, suffering catastrophic forgetting in lifelong settings, and being infeasible on resource-limited devices that cannot backpropagate.
Approach
Store per-domain affine-norm parameter deltas (domain vectors) on each device, share them across devices upon detected domain shifts, and either learn a reweighting over them plus new deltas via backprop (principal agents) or set reweighting weights from KL-based feature-statistic similarity without gradients (follower agents).
Key results
- On follower agents, CoLA (ETA) raises average ImageNet-C accuracy from 29.9% (NoAdapt) to 64.1%.
- With ETA+CoLA on principal agents, ImageNet-C lifelong average accuracy rises from 46.4% to 64.8% over 10 rounds.
- CoLA achieves up to 78.0x sample-efficiency speedup over ETA on ImageNet-C by leveraging shared knowledge.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CoLA to ETA | ETA · ImageNet-C level 5, lifelong 10-round adaptation | average accuracy (%): 46.4 -> 64.8 | Helped |
| Add CoLA to SAR | SAR · ImageNet-C level 5, lifelong 10-round adaptation | average accuracy (%): 60.5 -> 64.0 | Helped |
| Add CoLA to DeYO | DeYO · ImageNet-C level 5, lifelong 10-round adaptation | average accuracy (%): 10.9 -> 64.4 | Helped |
| Share and aggregate learned domain knowledge across principal devices | SAR · ImageNet-C level 5, collaborative adaptation | average accuracy (%): 58.6 -> 61.7 | Helped |
| Share and aggregate learned domain knowledge across principal devices | ETA · ImageNet-C level 5, collaborative adaptation | average accuracy (%): 61.2 -> 63.7 | Helped |
| Share and aggregate learned domain knowledge across principal devices | DeYO · ImageNet-C level 5, collaborative adaptation | average accuracy (%): 56.3 -> 62.6 | Helped |
| Aggregate shared ETA knowledge using forward-only similarity-based adaptation | Follower agent · ImageNet-C level 5, follower-device lifelong adaptation | average accuracy (%): 29.9 -> 64.1 | Helped |
| Add CoLA to ETA | ETA · ImageNet-C level 5, single-domain average | accuracy (%): 56.1 -> 59.3 | Helped |
| Add CoLA to SAR | SAR · ImageNet-C level 5, single-domain average | accuracy (%): 56.1 -> 58.1 | Helped |
| Add CoLA to DeYO | DeYO · ImageNet-C level 5, single-domain average | accuracy (%): 61.5 -> 62.5 | Helped |
| Add the learnable temperature Tl to knowledge reprogramming (ablation) | ETA + CoLA · ImageNet-C Gaussian, early adaptation with 960 samples | accuracy (%): 22.2% -> 53.2% | Helped |
| Remove the learnable temperature Tl (ablation) | ETA + CoLA · ImageNet-C Gaussian, early adaptation | accuracy (%): 51.7% accuracy with 640 samples versus 51.37% with 49,920 samples for ETA | Hurt |
Surprising findings
- Single-domain ETA+CoLA (64.4%) surpasses DeYO (64.1%) on the mild ImageNet-C scenario.
Limitations
- Relies on a fixed domain-shift detection threshold z and a fixed temperature Tf tuned without extensive search.
- Shared knowledge vectors accumulate over time, requiring pruning (e.g., max 32 weights in single-domain setting) to bound storage.
- Evaluated only on ImageNet-family corruption and shift benchmarks with ViT-Base and CLIP-RN50.
Experimental setup
Datasets: ImageNet-1k ImageNet-C ImageNet-R ImageNet-Sketch ImageNet-A ImageNet-V2
Benchmarks: ImageNet-C ImageNet-R ImageNet-Sketch ImageNet-A ImageNet-V2
Metrics: accuracy wall-clock time memory
Baselines: CoTTA ETA EATA SAR DeYO LAME T3A TPT FedTHE+ ATP
Builds on: ETA EATA SAR DeYO TPT
Closest papers by meaning
- BECoTTA: Input-dependent Online Blending of Experts for Continual Test-time Adaptation (ICML 2024 · similarity 0.85)
- ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation (ICLR 2024 · similarity 0.85)
- Class-aware Domain Knowledge Fusion and Fission for Continual Test-Time Adaptation (NeurIPS 2025 · similarity 0.84)
- Backpropagation-free Network for 3D Test-time Adaptation (CVPR 2024 · similarity 0.84)
- DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.84)
- SURGEON: Memory-Adaptive Fully Test-Time Adaptation via Dynamic Activation Sparsity (CVPR 2025 · similarity 0.84)
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.