Research map / Label-efficient and robust learning / Test-time adaptation
Feature-Based Instance Neighbor Discovery: Advanced Stable Test-Time Adaptation in Dynamic World
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 FIND, a test-time normalization method that partitions BN-layer features into instance-neighbor clusters (LFD), mixes test-specific and source BN statistics (FABN), and selectively applies partitioning per layer (S-FABN) to handle dynamic multi-distribution test batches.
Problem
Test-time adaptation degrades when test batches mix multiple distributions (dynamic scenarios), because global batch-norm statistics distort cross-distribution features.
Approach
Layer-wise feature disentanglement groups instance-level channel means by first-neighbor cosine similarity into connected components; each group's test-feature statistics are blended with source BN statistics (FABN), and a KL-based sensitivity score selects which layers are partitioned.
Key results
- FIND achieves 47.87% average accuracy across dynamic scenarios with ResNet-50 vs. 41.54% for the best prior baseline on Avg-All (alpha-BN 43.98%), with FIND* reaching 48.50%.
- On CrossMix with ResNet-50, FIND* reaches 47.19% average accuracy, versus 39.18% for alpha-BN and 39.47% for DeYO.
- FIND achieves 0.21 s latency and 1.44 GB memory on ImageNet-C with ResNet-50 on a V100, compared with 5.62 s for ViDA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FIND's feature partitioning and feature-aware normalization | ResNet-50 · CrossMix CIFAR10-C | accuracy (%): 68.85 -> 71.54 | Helped |
| Use FIND* selective layer partitioning instead of α-BN | ResNet-50 · CrossMix CIFAR100-C | accuracy (%): 33.22 -> 40.48{ ±0.1} | Helped |
| Use FIND* selective layer partitioning instead of α-BN | ResNet-50 · Random ImageNet-C | accuracy (%): 27.57 -> 30.62{ ±0.0} | Helped |
| Use FIND* selective layer partitioning instead of DeYO | ResNet-50 · Random CIFAR10-C | accuracy (%): 75.63 -> 73.68{ ±0.0} | Hurt |
| Use FIND* selective layer partitioning instead of DeYO | ResNet-50 · Shuffle CIFAR10-C | accuracy (%): 75.62 -> 73.60{ ±0.0} | Hurt |
| Add SBN statistics to TFN normalization (ablation) | FIND, ResNet-50 · CrossMix ImageNet-C | accuracy (%): 10.65 -> 29.21 | Helped |
| Add TFN target-distribution statistics to source-only SBN (ablation) | FIND, ResNet-50 · CrossMix ImageNet-C | accuracy (%): 25.64 -> 29.21 | Helped |
| Select layers for partitioning using sensitivity scores (ablation) | FIND, ResNet-50 · CrossMix ImageNet-C | accuracy (%): 29.21 -> 30.33 | Helped |
| Replace LFD with HDBSCAN clustering (ablation) | FIND · ImageNet-C | accuracy (%): 29.21 -> 24.66 | Hurt |
| Replace LFD with Agglomerative clustering (ablation) | FIND · ImageNet-C | accuracy (%): 29.21 -> 25.77 | Hurt |
| Replace LFD with K-means clustering (ablation) | FIND · ImageNet-C | accuracy (%): 29.21 -> 26.71 | Hurt |
| Set γ within 0-0.1 for selective partitioning (ablation) | FIND* · ImageNet-C and CIFAR10-C | accuracy (%): accuracy fluctuation remains below 1% | No clear effect |
Surprising findings
- Test-time fine-tuning (SAR) gives about -0.4% accuracy relative to test-time normalization in dynamic scenarios, suggesting normalization matters more than parameter updates there.
- Shallow and deep BN layers show lower domain-shift sensitivity than intermediate layers in ResNet-50, contrary to the common assumption.
Limitations
- Evaluated only on corruption-based benchmarks (CIFAR-C, ImageNet-C) with synthetic dynamic scenario constructions.
- Relies on BN layers, so applicability to architectures without BN (e.g., standard ViTs) is not shown.
- Cold-start phase and threshold gamma require tuning; FIND* and FIND show almost no accuracy gap.
Experimental setup
Datasets: CIFAR10-C CIFAR100-C ImageNet-C
Benchmarks: CIFAR10-C CIFAR100-C ImageNet-C
Metrics: accuracy latency memory
Baselines: SAR EATA DeYO TENT NOTE RoTTA ViDA TBN alpha-BN IABN
Builds on: TBN NOTE (IABN) alpha-BN FINCH
Closest papers by meaning
- TTVD: Towards a Geometric Framework for Test-Time Adaptation Based on Voronoi Diagram (ICLR 2025 · similarity 0.87)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.85)
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.85)
- Test-time Correlation Alignment (ICML 2025 · similarity 0.84)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.84)
- Improved Self-Training for Test-Time Adaptation (CVPR 2024 · similarity 0.83)
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.