Generalized Zero-Shot Learning With Multiple Graph Adaptive Generative Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generative Adversarial Networks (GANs) are used for generalized zero-shot learning (ZSL) to generate unseen image features from class embeddings.
• Existing GANs for ZSL often overlook the structural consistency between real and generated image features, potentially shifting distributions.
• This inconsistency can lead to generated features deviating from their true distributions, impacting ZSL performance.

Purpose of the Study:

• To propose a novel Multigraph Adaptive Generative Adversarial Network (MGA-GAN) for improved structure consistency in GAN-based ZSL.
• To align generator weights for better structural coherence between real and fake image features.
• To enhance the generation of discriminative features with improved structure consistency.

Main Methods:

• Developed MGA-GAN, a Wasserstein GAN incorporating a classification loss for generating discriminative features.
• Leveraged multigraph similarity structures of sliced real/fake feature samples to update generator weights within the local feature manifold.
• Employed correlation graphs of global real/fake features to ensure structure correlation across the entire feature manifold.

Main Results:

• MGA-GAN demonstrated superior performance compared to state-of-the-art methods on four benchmark datasets.
• The proposed method effectively improves structure consistency between real and generated image features.
• Evaluations confirmed the effectiveness of leveraging multigraph structures for generator weight updates.

Conclusions:

• MGA-GAN offers a significant advancement in GAN-based generalized zero-shot learning by addressing feature distribution structure consistency.
• The approach successfully generates more accurate and discriminative features for unseen classes.
• The findings highlight the importance of structural alignment in GANs for robust ZSL.