Reproduce “A Review of Generalized Zero-Shot Learning Methods” on OpenResearch

About this paper

Farhad Pourpanah, Moloud Abdar, Yuxuan Luo, Xinlei Zhou, Ran Wang, Chee Peng Lim, Xi-Zhao Wang, Q. M. Jonathan Wu

This paper offers the first comprehensive and in-depth review of Generalized Zero-Shot Learning (GZSL) methods, systematically categorizing them into embedding-based and generative-based approaches while detailing their core challenges and outlining future research directions to enable AI systems to classify both known and novel object categories simultaneously.

Key insights

• GZSL methods face fundamental challenges including the hubness problem in high-dimensional spaces, a projection domain shift between seen and unseen class distributions, and an inherent bias towards better-represented seen classes.
• Semantic information, whether human-defined attributes or word vectors, is crucial for transferring knowledge to unseen classes, but its quality and generation methods are key areas for improvement.
• Latent space embedding and utilizing transductive settings (where unlabeled unseen data is available during training) show promise in mitigating the hubness problem and improving the balance of seen and unseen class accuracies.

Summary by alphaXiv.