Research map / Label-efficient and robust learning / Recognition and long-tail
A Study on PAVE Specification for Learnware
Hao-Yu Shi, Zhi-Hao Tan, Zi-Chen Zhao, Yang Yu, Zhi-Hua Zhou
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 |
|---|---|---|---|
| fit model predictions p(y_hat|x) instead of data distribution p(y|x) for model vectors (ablation) | PAVE · CV with corrupted learnwares (average) | accuracy: 0.745 -> 0.887 | Improved |
| fit model predictions instead of data distribution (ablation) | PAVE · CIFAR10 (corrupted learnwares) | accuracy: 0.894 -> 0.890 | No clear effect |
| fit model predictions instead of data distribution (ablation) | PAVE · GTSRB (corrupted learnwares) | accuracy: 0.656 -> 0.987 | Improved |
| fit model predictions instead of data distribution (ablation) | PAVE · EuroSAT (corrupted learnwares) | accuracy: 0.742 -> 0.991 | Improved |
| fit model predictions instead of data distribution (ablation) | PAVE · SVHN (corrupted learnwares) | accuracy: 0.643 -> 0.970 | Improved |
| identify learnware with PAVE instead of RKME, then fine-tune | PAVE · NLP beyond original functionality (average) | average score: 0.659 -> 0.709 | Improved |
| identify learnware with PAVE instead of random selection | PAVE · NLP beyond original functionality (average) | average score: 0.668 -> 0.709 | Improved |
| fine-tune identified learnware instead of fine-tuning RoBERTa-L | PAVE · NLP beyond original functionality (average) | average score: 0.699 -> 0.709 | Improved |
| PAVE instead of RKME on QQP accuracy | PAVE · QQP (beyond original functionality) | Acc: 0.696 -> 0.631 | Hurt |
| PAVE instead of RKME on RTE accuracy | PAVE · RTE (beyond original functionality) | Acc: 0.635 -> 0.560 | Hurt |
| identify learnware with PAVE instead of random on corrupted CV repository | PAVE · CV with corrupted learnwares (average) | average accuracy: 0.580 -> 0.887 | Improved |
| identify learnware with PAVE instead of fine-tuning ViT-L-14 | PAVE · CV with corrupted learnwares (average) | average accuracy: 0.733 -> 0.887 | Improved |
Closest papers by meaning
- On the Ability of Developers' Training Data Preservation of Learnware (NeurIPS 2024 · similarity 0.86)
- Towards Making Learnware Specification and Market Evolvable (AAAI 2024 · similarity 0.86)
- Learnware Specification via Label-Aware Neural Embedding (AAAI 2025 · similarity 0.86)
- Learnware Specification via Dual Alignment (ICML 2025 · similarity 0.86)
- Handling Learnwares from Heterogeneous Feature Spaces with Explicit Label Exploitation (NeurIPS 2024 · similarity 0.86)
- Domain2Vec: Vectorizing Datasets to Find the Optimal Data Mixture without Training (ICML 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.
