Research map / Label-efficient and robust learning / Test-time adaptation
Better Aggregation in Test-Time Augmentation
Divya Shanmugam, Davis Blalock, Guha Balakrishnan, John Guttag
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace mean with learned per-augmentation/class weights | TTA aggregation, MobileNetV2 · Flowers102 (standard policy) | Top-1 accuracy: 90.47 -> 92.62 | Improved |
| replace mean with learned weights | TTA aggregation, InceptionV3 · ImageNet (standard policy) | Top-1 accuracy: 72.45 -> 72.79 | Improved |
| replace mean with learned weights | TTA aggregation, MobileNetV2 · ImageNet (standard policy) | Top-1 accuracy: 72.69 -> 72.43 | Hurt |
| replace mean with learned weights | TTA aggregation, CNN-7 · CIFAR100 (expanded policy) | Top-1 accuracy: 73.41 -> 73.18 | Hurt |
| replace mean with learned weights | TTA aggregation, CNN-5 · STL10 (standard policy) | Top-1 accuracy: 78.58 -> 78.52 | No clear effect |
| replace mean with learned weights | TTA aggregation, ResNet-50 · Flowers102 (expanded policy) | Top-1 accuracy: 92.48 -> 93.29 | Improved |
| use mean-logit TTA instead of no TTA | ResNet-50 · ImageNet (standard policy) | Top-1 accuracy: 75.78 -> 76.91 | Improved |
| use mean-logit TTA instead of no TTA with expanded policy | CNN-7 · CIFAR100 (expanded policy) | Top-1 accuracy: 74.37 -> 73.41 | Hurt |
| aggregate with Max instead of Mean under expanded policy (ablation) | ResNet-18 · Flowers102 (expanded policy) | Top-1 accuracy: 89.47 -> 83.30 | Hurt |
| use greedy policy search (GPS) instead of Mean (ablation) | MobileNetV2 · Flowers102 (standard policy) | Top-1 accuracy: 90.47 -> 88.28 | Hurt |
| small model with learned TTA vs larger model without TTA | MobileNetV2 vs ResNet-50 · Flowers102 (standard policy) | Top-1 accuracy: 91.72 -> 92.62 | Improved |
| expanded instead of standard TTA policy for learned aggregation (ablation) | learned TTA aggregation, MobileNetV2 · Flowers102 | Top-1 accuracy: 92.62 -> 92.49 | No clear effect |
Closest papers by meaning
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.85)
- Self-Bootstrapping for Versatile Test-Time Adaptation (ICML 2025 · similarity 0.85)
- Robust Test-Time Adaptation for Zero-Shot Prompt Tuning (AAAI 2024 · similarity 0.84)
- DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · similarity 0.84)
- Test-Time Adaptive Object Detection with Foundation Model (NeurIPS 2025 · similarity 0.84)
- Test-time Augmentation Improves Efficiency in Conformal Prediction (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.
