Reproduce “Absolute 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
Rui Gao, Fan Wan, Daniel Organisciak, Jiyao Pu, Junyan Wang, Haoran Duan, Peng Zhang, Xingsong Hou and 1 more
A novel Absolute Zero-Shot Learning (AZSL) paradigm enables training models for zero-shot classification without access to any real data, using a pre-trained teacher model to provide supervision signals to a generator and student network. This framework achieved accuracy close to 70% for unseen classes on the AWA1 dataset, demonstrating competitive performance with privacy preservation.
Key insights
- It is possible to train Zero-Shot Learning models entirely without real data by leveraging knowledge transfer from a pre-trained teacher model.
- Synthetic feature generation, guided by a teacher model, can effectively mimic the distribution of real data for downstream classification tasks.
- The framework can operate in both white-box (teacher model architecture and weights accessible) and black-box (only teacher outputs accessible) scenarios, balancing performance against teacher model security.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.