Research map / Vision-language and multimodal / CLIP and zero-shot
LOGICZSL: Exploring Logic-induced Representation for 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.
Proposes LOGICZSL, which uses GPT-4-generated relational knowledge (object-induction, object-extension, attribute-extension trees) turned into fuzzy first-order logic losses that plug into existing CLIP-based CZSL methods (CSP, Troika).
Problem
CLIP-based compositional zero-shot learning ignores semantic relationships among attributes, objects, and their compositions (object hierarchy, inter-class similarity, attribute polysemy), limiting generalization to unseen attribute-object compositions.
Approach
Query GPT-4 to build primitive concept trees (object groups, object descriptions, attribute synonyms/extensions), express them as first-order logic rules, relax them with fuzzy logic and the generalized mean quantifiers, and add the resulting differentiable logic losses to standard CZSL losses so any CLIP-based CZSL method can be augmented.
Key results
- Logic-Troika reaches 15.3 AUC on CGQA closed world vs 12.4 for Troika.
- Logic-CSP improves UT-Zappos50K closed-world AUC from 33.0 to 38.2.
- Logic-Troika improves MIT-States closed-world AUC from 22.1 to 23.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LOGICZSL logic-induced learning to Troika | Troika · CGQA close world | AUC: 12.4 → 15.3 | Helped |
| Add LOGICZSL logic-induced learning to Troika | Troika · CGQA open world | AUC: 2.7 → 3.4 | Helped |
| Add LOGICZSL logic-induced learning to Troika | Troika · UT-Zappos50K close world | AUC: 41.7 → 45.8 | Helped |
| Add LOGICZSL logic-induced learning to Troika | Troika · UT-Zappos50K open world | AUC: 33.0 → 36.2 | Helped |
| Add LOGICZSL logic-induced learning to Troika | Troika · MIT-States close world | AUC: 22.1 → 23.4 | Helped |
| Add LOGICZSL logic-induced learning to Troika | Troika · MIT-States open world | AUC: 7.2 → 8.7 | Helped |
| Add oi-logic (ablation) | Troika · UT-Zappos50K close world | AUC: 41.7 → 45.6 | Helped |
| Add oi-logic (ablation) | Troika · UT-Zappos50K open world | AUC: 33.0 → 35.0 | Helped |
| Add oe-logic (ablation) | Troika · UT-Zappos50K close world | AUC: 41.7 → 45.1 | Helped |
| Add ae-logic (ablation) | Troika · UT-Zappos50K open world | AUC: 33.0 → 35.2 | Helped |
| Set aggregation coefficient q to 3.0 rather than q=1.0 (ablation) | Logic-Troika · UT-Zappos50K close world | AUC: 45.1 → 45.5 | Helped |
| Set aggregation coefficient q to 3.0 rather than q=5.0 (ablation) | Logic-Troika · UT-Zappos50K open world | AUC: 34.2 → 35.3 | Helped |
Limitations
- Gains on some settings are modest (e.g., Logic-CSP CGQA open-world AUC 1.2 to 1.4)
- Logic loss weight needs scheduling (activated after several epochs) to help
Experimental setup
Datasets: CGQA UT-Zappos50K MIT-States
Benchmarks: CGQA UT-Zappos50K MIT-States
Metrics: AUC HM Seen accuracy Unseen accuracy
Baselines: CoOp Prod HPL PCVL DFSP PLID CDS-CZSL CAILA CSP Troika
Builds on: CSP Troika CLIP fuzzy logic (Badreddine-style logic tensor relaxations)
Closest papers by meaning
- Beyond Seen Primitive Concepts and Attribute-Object Compositional Learning (CVPR 2024 · similarity 0.92)
- A Dynamic Learning Method towards Realistic Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.89)
- Retrieval-Augmented Primitive Representations for Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.89)
- Revealing the Proximate Long-Tail Distribution in Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.88)
- Context-based and Diversity-driven Specificity in Compositional Zero-Shot Learning (CVPR 2024 · similarity 0.88)
- TOMCAT: Test-time Comprehensive Knowledge Accumulation for Compositional Zero-Shot Learning (NeurIPS 2025 · similarity 0.87)
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.