VMAN: A Virtual Mainstay Alignment Network for Transductive Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Conventional zero-shot learning (ZSL) methods struggle with adapting learned weights to unseen images.
• Transductive zero-shot learning (TZSL) aims to improve ZSL by utilizing unlabeled unseen images during training.
• Existing TZSL methods often have weight embeddings dominated by seen classes, limiting generalization.

Purpose of the Study:

• To propose the Virtual Mainstay Alignment Network (VMAN) for effective knowledge transfer in transductive zero-shot learning.
• To address the limitation of weight embeddings being biased towards seen classes in TZSL.
• To enhance the adaptation of model weights for better performance on unseen images.

Main Methods:

• VMAN utilizes a tied encoder-decoder network, learning a single set of linear mapping weights.
• Introduces virtual mainstay (VM) samples for each seen class to serve as novel training data.
• Incorporates a weighted reconstruction scheme and instance-category matching regularization for improved weight alignment.

Main Results:

• VMAN effectively aligns embedding weights between seen and unseen classes.
• The generation of VM samples prevents weight bias towards seen images.
• Achieves superior performance in (Generalized) TZSL settings across four benchmark datasets.

Conclusions:

• VMAN offers a novel and effective approach for transductive zero-shot learning.
• The proposed methods significantly improve the adaptation of models to unseen classes.
• VMAN demonstrates state-of-the-art results in TZSL tasks.