Reproduce “Data-Free Generalized Zero-Shot Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Bowen Tang, Long Yan, Jing Zhang, Qian Yu, Lu Sheng, Dong Xu

Researchers from Beihang University and The University of Hong Kong developed a three-stage framework for Data-Free Generalized Zero-Shot Learning (DFZSL), enabling classification of both known and novel classes without access to any real images. Their method, which includes virtual data generation and feature-language prompt tuning, improved harmonic mean accuracy by up to 10.8% on GZSL benchmarks and showed better generalization on 11 diverse image classification datasets.

Key insights

  • L2-normed CLIP image features can be effectively approximated by a von Mises-Fisher distribution, enabling the recovery of virtual data distributions even without real images.
  • A data-driven, multi-modal feature-language prompt tuning (FLPT) method effectively bridges the modality gap between image and text features from CLIP for fine-grained ZSL tasks without requiring real data.
  • Combining virtual data recovery, multi-modal prompt tuning, and conditional generative modeling provides a robust and practical pipeline for Generalized Zero-Shot Learning under severe data-free constraints.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.