Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Data-Free Generalized 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 data-free zero-shot learning, recovering virtual CLIP base-class features via a vMF distribution, aligning them with text via feature-language prompt tuning, and training a conditional generator for new classes.
Problem
Zero-shot learning normally requires real base-class images and manual attributes, which are unavailable under copyright or privacy restrictions; the paper targets zero-shot classification using only a CLIP-based pre-trained base-class classifier.
Approach
Model CLIP image features of each base class as samples from a von Mises-Fisher distribution centered on the classifier prototypes to recover virtual features, align them with CLIP text features through learnable feature-language prompts and a shift term, then train a conditional generator (e.g., TF-VAEGAN) on the aligned features to synthesize unseen-class features.
Key results
- On generalized ZSL, FLPT+TF-VAEGAN in the black-box setting reaches 93.5 H on AWA2 and 82.6 H on APY, improving over SHIP+CoOp's 89.0 H on AWA2 by about 4.5 points.
- On base-to-new generalization averaged over 11 datasets, FLPT+TFVAEGAN reaches 79.71 H versus 76.73 H for SHIP+CoOp.
- Virtual base-class data gives performance close to real data on AWA2, APY and FLO, per the ablation comparing real and recovered features.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FLPT+TF-VAEGAN instead of the strongest listed baseline | Data-free GZSL · AWA2 | H: 89.0 -> 93.5 | Helped |
| Use FLPT+TF-VAEGAN instead of the strongest listed baseline | Data-free GZSL · APY | H: 80.5 -> 82.6 | Helped |
| Use FLPT+TF-VAEGAN instead of the strongest listed baseline | Data-free GZSL · CUB | H: 68.1 -> 63.4 | Hurt |
| Use FLPT+TF-VAEGAN instead of the strongest listed baseline | Data-free GZSL · SUN | H: 43.0 -> 63.2 | Helped |
| Use FLPT+TF-VAEGAN instead of the strongest listed baseline | Data-free GZSL · FLO | H: 71.7 -> 83.2 | Helped |
| Add FLPT prompt tuning instead of hand-crafted prompts (ablation) | Data-free GZSL · AWA2 | H: 90.57 -> 93.24 | Helped |
| Add FLPT prompt tuning instead of CoOp (ablation) | Data-free GZSL · CUB | H: 51.82 -> 57.50 | Helped |
| Add TF-VAEGAN generation to FLPT (ablation) | Data-free GZSL · AWA2 | H: 93.24 -> 93.51 | Helped |
| Add ZLAP generation to FLPT (ablation) | Data-free GZSL · CUB | H: 57.50 -> 63.14 | Helped |
| Add SDGZSL generation to FLPT (ablation) | Data-free GZSL · FLO | H: 71.57 -> 82.26 | Helped |
| Replace recovered virtual base data with real base data (ablation) | FLPT+TF-VAEGAN · AWA2 | H: 93.51 -> 93.71 | No clear effect |
| Replace recovered virtual base data with real base data (ablation) | FLPT+TF-VAEGAN · CUB | H: 63.39 -> 65.02 | Helped |
Surprising findings
- The data-free setting with recovered virtual features is competitive with, and on some benchmarks (e.g., AWA2 and APY) close to, training on real base data.
Limitations
- Performance gap to real-data training remains on fine-grained datasets such as CUB and SUN.
- Relies on a CLIP-based base classifier being available; the framework's generality to other vision-language models is claimed but not evaluated.
- Black-box class prototype alignment depends on the server returning prediction scores.
Experimental setup
Datasets: AWA2 APY CUB SUN FLO ImageNet Caltech101 OxfordPets StanfordCars Flowers102 Food101 FGVCAircraft
Benchmarks: AWA2 APY CUB SUN FLO ImageNet Caltech101 OxfordPets StanfordCars Flowers102 Food101 FGVCAircraft
Metrics: top-1 accuracy harmonic mean
Baselines: f-CLSWGAN Cycle-WGAN LisGAN TCN f-VAEGAN TF-VAEGAN GCM-CF HSVA MSDN AZSL SHIP+CoOp CoOp
Builds on: CLIP CoOp CoCoOp TF-VAEGAN AZSL
Closest papers by meaning
- Semantic Feature Extraction for Generalized Zero-Shot Learning (AAAI 2022 · similarity 0.92)
- GenZSL: Generative Zero-Shot Learning Via Inductive Variational Autoencoder (ICML 2025 · similarity 0.91)
- Data Distribution Distilled Generative Model for Generalized Zero-Shot Recognition (AAAI 2024 · similarity 0.91)
- Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names (CVPR 2024 · similarity 0.91)
- TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · similarity 0.90)
- ZeroDiff: Solidified Visual-semantic Correlation in Zero-Shot Learning (ICLR 2025 · similarity 0.89)
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.