# Bi-directional Distribution Alignment for Transductive Zero-Shot Learning

**Authors:** Zhicai Wang, Yanbin Hao, Tingting Mu, Ouxiang Li, Shuo Wang, Xiangnan He  
**arXiv:** [2303.08698](https://arxiv.org/abs/2303.08698)  
**Date:** 2023-03-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.08698v1/

## Key points

- Zero-shot learning (ZSL) has a problem of domain shift where true and learned data distributions for unseen classes do not match.
- Transductive ZSL (TZSL) attempts to alleviate this issue by allowing the use of unlabelled examples from unseen classes, but there is still a high level of distribution shift.
- A team of researchers led by Zhicai Wang proposed a novel TZSL model called Bi-VAEGAN that significantly improves distribution alignment between visual and auxiliary spaces.
- Bi-VAEGAN's key design proposals include bi-directional distribution alignment, L_2-norm based feature normalization approach, and a more sophisticated estimation approach for unseen class priors.
- In benchmark evaluations using four datasets, Bi-VAEGAN achieved new state-of-the-art results under both standard and generalized TZSL settings.
- Bi-VAEGAN provides a more robust framework for TZSL that can better handle domain shift issues.
- The code for Bi-VAEGAN is available on GitHub at https://github.com/Zhicaiwww/Bi-VAEGAN.

## Summary

The field of zero-shot learning (ZSL) has long been plagued by the problem of domain shift, where the true and learned data distributions for unseen classes do not match. While transductive ZSL (TZSL) has attempted to alleviate this issue by allowing the use of unlabelled examples from unseen classes, there remains a high level of distribution shift. To address this challenge, a team of researchers led by Zhicai Wang have proposed a novel TZSL model called Bi-VAEGAN that significantly improves distribution alignment between visual and auxiliary spaces. Bi-VAEGAN's key design proposals include bi-directional distribution alignment, a simple yet effective L_2-norm based feature normalization approach, and a more sophisticated estimation approach for unseen class priors. In benchmark evaluations using four datasets, Bi-VAEGAN achieved new state-of-the-art results under both standard and generalized TZSL settings. This research is significant because it offers an innovative solution to one of the most pressing challenges in ZSL. By improving distribution alignment between visual and auxiliary spaces, Bi-VAEGAN provides a more robust framework for TZSL that can better handle domain shift issues. The code for Bi-VAEGAN is available on GitHub at https://github.com/Zhicaiwww/Bi-VAEGAN , making it accessible to other researchers who want to build on this work.