Research map / Label-efficient and robust learning / Test-time adaptation
Deep Nets with Subsampling Layers Unwittingly Discard Useful Activations at Test-Time
Chiao-An Yang, Ziwei Liu, Raymond Yeh
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 |
|---|---|---|---|
| add search-and-aggregate of discarded subsampling activations to existing TTA methods | GPS, ClassTTA, AugTTA · ImageNet | top-1 accuracy: ClassTTA ResNet18 B=150: 66.40 -> 70.37 | Improved |
| add the method on top of TTA baselines | GPS, ClassTTA, AugTTA · Flowers102 | top-1 accuracy: GPS ResNet18 B=30: 89.04 -> 88.93; AugTTA B=100: 88.86 -> 89.55 | Mixed |
| increase total test-time budget for baseline TTA methods | ClassTTA · ImageNet | top-1 accuracy: ResNet18: 69.09 -> 66.40 | Hurt |
| add spatial alignment before aggregation (ablation) | learned aggregation · ImageNet (own val split) | top-1 accuracy: 79.52 -> 79.88 | Improved |
| use learned aggregation instead of simple average (ablation) | aggregation · ImageNet (own val split) | top-1 accuracy: 79.38 -> 79.88 | Improved |
| omit first and last layers from search space (ablation) | search · ImageNet (own val split) | top-1 accuracy: 79.64 -> 79.88 | Improved |
| use learned attention criterion instead of random expansion (ablation) | search · ImageNet (own val split) | top-1 accuracy: 79.80 -> 79.88 | Improved |
| use entropy criterion (training-free) instead of learned criterion (ablation) | search · ImageNet (own val split) | top-1 accuracy: 79.88 -> 79.86 | No clear effect |
| increase B_ours beyond about 15-20 (ablation) | search budget · ImageNet | top-1 accuracy | No clear effect |
| add non-learned method to segmentation models | FCN, DeepLab, SegFormer · Cityscapes | mIoU: 76.54 -> 77.05 (SegFormer, B=10) | Improved |
| combine method with horizontal flip TTA | segmentation models · ADE20K | mIoU: 35.94 -> 36.31 (flip) -> 36.50 (flip + ours) | Improved |
Closest papers by meaning
- BlockDrop: Dynamic Inference Paths in Residual Networks (CVPR 2018 · similarity 0.86)
- How to Start Training: The Effect of Initialization and Architecture (NeurIPS 2018 · similarity 0.86)
- Residual Alignment: Uncovering the Mechanisms of Residual Networks (NeurIPS 2023 · similarity 0.85)
- DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging (NeurIPS 2026 · similarity 0.84)
- Not All Diffusion Model Activations Have Been Evaluated as Discriminative Features (NeurIPS 2024 · similarity 0.84)
- Emphasizing Discriminative Features for Dataset Distillation in Complex Scenarios (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.
