Zero-VAE-GAN: Generating Unseen Features for Generalized and Transductive Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) faces challenges due to the absence of unseen class data during training.
• Existing ZSL methods often suffer from a domain shift problem, biasing models towards seen classes and limiting performance.

Purpose of the Study:

• To address the domain shift problem in ZSL by converting it into a supervised learning task through feature generation for unseen classes.
• To propose a novel generative model, Zero-VAE-GAN, for creating high-quality unseen features.

Main Methods:

• A joint generative model, Zero-VAE-GAN, coupling Variational Autoencoder (VAE) and Generative Adversarial Network (GAN), is developed to generate unseen class features.
• An adversarial categorization network is integrated to enhance class-level discriminability.
• Two self-training strategies are introduced for transductive ZSL, augmenting unlabeled unseen features.

Main Results:

• The proposed Zero-VAE-GAN model demonstrates superior performance over state-of-the-art methods on conventional and generalized ZSL tasks across five benchmarks and a large-scale dataset.
• The transductive extension using self-training strategies further enhances performance, effectively addressing the domain shift problem.

Conclusions:

• Zero-VAE-GAN offers a robust solution for ZSL by generating high-quality unseen features and mitigating domain shift.
• The proposed self-training strategies prove effective in improving ZSL performance in a transductive setting.