Research map / Vision-language and multimodal / CLIP and zero-shot
Global and Local Entailment Learning for Natural World Imagery
Srikumar Sastry, Aayush Dhakal, Eric Xing, Subash Khanal, Nathan Jacobs
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 global entailment loss L_GE to local entailment loss (ablation) | RCME · iNaturalist-2021 zero-shot classification | average accuracy: 41.50 -> 62.97 | Improved |
| replace prior preservation loss with cross-modal alignment loss (ablation) | RCME · iNaturalist-2021 zero-shot classification | average accuracy: 62.97 -> 65.09 | Improved |
| add cross-modal alignment loss to local entailment loss alone (ablation) | L_LE · iNaturalist-2021 zero-shot classification | average accuracy: 41.50 -> 43.65 | Improved |
| add hard negative mining (ablation) | RCME · iNaturalist-2021 zero-shot classification | average accuracy: 62.22 -> 65.09 | Improved |
| add hard negative mining (ablation) | RCME · BioCLIP-Rare zero-shot classification | average accuracy: 59.12 -> 62.64 | Improved |
| use RCME instead of Radial Embeddings for ordering of taxonomic labels | RCME · iNaturalist-2021 hierarchical retrieval | Kendall's tau_d: 0.521 -> 0.993 | Improved |
| use RCME instead of ATMG for zero-shot classification | RCME · iNaturalist-2021 zero-shot classification | average accuracy: 61.89 -> 65.09 | Improved |
| train from OpenCLIP initialization instead of fine-tuning from BioCLIP | RCME · iNaturalist-2021 zero-shot classification | average accuracy: 57.71 -> 65.09 | Improved |
| use RCME for image-to-image retrieval versus ATMG | RCME · iNaturalist-2021 image-to-image retrieval | average R@1: 84.08 -> 86.88 | Improved |
| apply RCME objective on HierarCaps with CLIP ViT-B/16 versus Radial Emb. | RCME · HierarCaps | F1: 0.229 -> 0.241 | Improved |
| vary margin alpha of global entailment loss from 0 to pi/2 (ablation) | RCME · HierarCaps | F1: 0.226 -> 0.241 | Improved |
| reduce number of retrieval steps from 50 to 10 (ablation) | RCME · iNaturalist-2021 hierarchical retrieval | precision and recall: 0.458 -> 0.527 precision; 0.572 -> 0.472 recall | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BioCLIP — BioCLIP: A Vision Foundation Model for the Tree of Life (CVPR 2024) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Compositional Entailment Learning for Hyperbolic Vision-Language Models (ICLR 2025 · similarity 0.89)
- Natural Language Inference Improves Compositionality in Vision-Language Models (ICLR 2025 · similarity 0.87)
- Defeasible Visual Entailment: Benchmark, Evaluator, and Reward-Driven Optimization (AAAI 2025 · similarity 0.87)
- Language Model as Visual Explainer (NeurIPS 2024 · similarity 0.86)
- Progressive Multi-granular Alignments for Grounded Reasoning in Large Vision-Language Models (AAAI 2025 · similarity 0.86)
- Aligning Forest and Trees in Images & Long Captions for Visually Grounded Understanding (NeurIPS 2026 · similarity 0.86)
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.
