Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Revealing the Proximate Long-Tail Distribution in Compositional Zero-Shot Learning
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.
Reframes compositional zero-shot learning as a proximate long-tailed problem and applies logit adjustment using an attribute prior estimated from state/object classifier posteriors, improving CZSL without extra parameters.
Problem
Compositional zero-shot learning suffers from visual bias caused by state-object interactions, which blurs visual features and hinders learning distinguishable class prototypes; the results approximate a long-tailed distribution.
Approach
Treats visual bias from state-object compositions as an "attribute imbalance" that acts as an approximate class prior, estimated from the posteriors of independent state and object classifiers, and incorporates it via logit adjustment during training and inference (ProLT), derived from a mutual-information and HM lower-bound analysis.
Key results
- On MIT-States with ResNet-18, ProLT reaches AUC 6.0 and HM 19.3, surpassing CANet's AUC by 0.6 points.
- On UT-Zappos with ResNet-18, ProLT achieves HM 49.3, outperforming CANet by 2.0 points.
- With CLIP ViT-L/14 backbone, ProLT achieves HM 38.2 on MIT-States and 49.4 on UT-Zappos, claimed as state-of-the-art among compared CLIP-based methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ProLT instead of OADis | ResNet-18 CZSL · MIT-States | AUC: 5.9 -> 6.0 | Helped |
| Use ProLT instead of CANet | ResNet-18 CZSL · UT-Zappos | HM: 47.3 -> 49.3 | Helped |
| Use ProLT instead of CANet | ResNet-18 CZSL · C-GQA | HM: 14.5 -> 14.4 | Hurt |
| Use ProLT instead of DFSP | CLIP CZSL · UT-Zappos | HM: 47.2 -> 49.4 | Helped |
| Add the attribute prior rather than use no prior (ablation) | ResNet-18 ProLT with FC prototype learner · UT-Zappos | AUC: 32.3 -> 33.4 | Helped |
| Replace the dataset class prior with the attribute prior (ablation) | ResNet-18 ProLT with FC prototype learner · UT-Zappos | AUC: 32.0 -> 33.4 | Helped |
| Set η to 0 rather than use the full adjustment (ablation) | ResNet-18 ProLT with FC prototype learner · UT-Zappos | AUC: 33.4 -> 28.9 | Hurt |
| Remove the inference prior (ablation) | ResNet-18 ProLT with FC prototype learner · UT-Zappos | AUC: 33.4 -> 32.7 | Hurt |
| Add the attribute prior rather than use no prior (ablation) | CLIP ProLT · UT-Zappos | AUC: 33.6 -> 36.1 | Helped |
| Replace the dataset class prior with the attribute prior (ablation) | CLIP ProLT · UT-Zappos | AUC: 34.2 -> 36.1 | Helped |
| Set η to 0 rather than use the full adjustment (ablation) | CLIP ProLT · UT-Zappos | AUC: 36.1 -> 31.7 | Hurt |
| Remove the inference prior (ablation) | CLIP ProLT · UT-Zappos | AUC: 36.1 -> 35.1 | Hurt |
Surprising findings
- Using a class prior from dataset counts gives smaller gains than the attribute prior estimated from model posteriors, suggesting visual bias matters more than class frequency.
- Setting eta to 0 (removing logit adjustment in training) causes the largest performance drop in ablations.
Limitations
- Performs slightly below state-of-the-art on C-GQA (HM 14.4 vs CANet's 14.5 and lower AUC gap noted).
- Struggles with fine-grained distinctions such as among boot types on UT-Zappos.
- Unseen attribute prior is approximated via importance sampling with high variance when only one sample is available at test time.
Experimental setup
Datasets: MIT-States UT-Zappos C-GQA ImageNet
Benchmarks: MIT-States UT-Zappos C-GQA
Metrics: AUC HM best seen accuracy best unseen accuracy state accuracy object accuracy
Baselines: CompCos CGE SCEN Co-CGE OADis DECA CANet CSP DFSP
Builds on: logit adjustment CGE GZSL
Closest papers by meaning
- A Dynamic Learning Method towards Realistic Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.90)
- Beyond Seen Primitive Concepts and Attribute-Object Compositional Learning (CVPR 2024 · similarity 0.90)
- Context-based and Diversity-driven Specificity in Compositional Zero-Shot Learning (CVPR 2024 · similarity 0.89)
- Retrieval-Augmented Primitive Representations for Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.89)
- Beyond Image Classification: A Video Benchmark and Dual-Branch Hybrid Discrimination Framework for Compositional Zero-Shot Learning (CVPR 2025 · similarity 0.89)
- Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations (NeurIPS 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.