Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Adaptive Confidence Smoothing for Generalized Zero-Shot Learning
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 COSMO confidence-based gating and adaptive smoothing to LAGO (ablation) | LAGO · AWA, SUN, CUB GZSL test | Acc_H: 58.3 -> 63.6; 35.1 -> 41.0; 44.6 -> 50.2 | Helped |
| Add adaptive confidence smoothing instead of no smoothing (ablation) | COSMO+LAGO · AWA, SUN, CUB GZSL-Val | Acc_H: 52.9 -> 54.3; 38.4 -> 41.4; 43.4 -> 45.7 | Helped |
| Use adaptive rather than constant smoothing weight λ (ablation) | COSMO+LAGO · AWA, SUN, CUB GZSL-Val | Acc_H: 53 -> 54.3; 38.6 -> 41.4; 43.6 -> 45.7 | Helped |
| Add confidence-based gating rather than independent hard gating (ablation) | COSMO+LAGO · AWA, SUN, CUB test | Acc_H: 58.3 -> 64.3; 35.1 -> 39.7; 44.6 -> 49.1 | Helped |
| Add adaptive smoothing to confidence-based gating (ablation) | COSMO+LAGO · AWA, SUN, CUB test | Acc_H: 64.3 -> 63.6; 39.7 -> 41.0; 49.1 -> 50.2 | Mixed |
| Use temperature-scaled CB-Gating-3 (T=3) rather than CB-Gating-1 (T=1) (ablation) | COSMO+LAGO gating module · AWA, SUN, CUB GZSL-Val | AUC: 88.9 -> 92.5; 75.5 -> 77.7; 81.7 -> 82.0 | Mixed |
| Use confidence-based gating instead of max-softmax gating (ablation) | COSMO+LAGO gating module · AWA, SUN, CUB GZSL-Val | AUC: 86.7 -> 92.5; 60.9 -> 77.7; 74.1 -> 82.0 | Helped |
| Remove predictions from the zero-shot expert from the confidence gate (ablation) | CB-Gating-3 · AWA, SUN, CUB GZSL-Val | AUC: 92.5 -> 88.8; 77.7 -> 61; 82.0 -> 74.2 | Mixed |
| Jointly train modules while excluding unseen validation classes from seen-expert training (ablation) | COSMO on CUB · CUB GZSL | Acc_H: 50.2% -> 26.3% | Hurt |
| Jointly train modules using all labeled samples as seen, leaving no unseen-class training samples (ablation) | COSMO on CUB · CUB GZSL | Acc_H: Acc_ts = 0.1%, Acc_tr = 72.8%, Acc_H = 1.9% | Hurt |
| Apply input perturbation in ODIN and Ensemble OOD detection | OOD detection on ZSL benchmarks · CUB, SUN, AWA | OOD metrics | Hurt |
| Train an Ensemble OOD model on held-out classes | OOD detection · Unseen-Val | OOD detection | Hurt |
Closest papers by meaning
- Unseen Classes at a Later Time? No Problem (CVPR 2022 · similarity 0.91)
- Data Distribution Distilled Generative Model for Generalized Zero-Shot Recognition (AAAI 2024 · similarity 0.90)
- Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names (CVPR 2024 · similarity 0.89)
- Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations (NeurIPS 2024 · similarity 0.88)
- Data-Free Generalized Zero-Shot Learning (AAAI 2024 · similarity 0.88)
- Zero-Shot Robustification of Zero-Shot Models (ICLR 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.