Generalized zero-shot domain adaptation via coupled conditional variational autoencoders

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Domain adaptation leverages source data for target domain learning when target data is scarce.
• Existing methods often assume all target classes are present, overlooking scenarios with partial class availability.
• This limitation hinders effective model training in real-world applications with evolving or incomplete datasets.

Purpose of the Study:

• To address the challenge of domain adaptation with a subset of available classes in the target domain.
• To propose a novel framework that integrates domain adaptation and generalized zero-shot learning.
• To develop a method capable of generating synthetic features for unseen classes in the target domain.

Main Methods:

• Formulated the problem within a generalized zero-shot learning framework, using source domain data as semantic representations.
• Introduced a Coupled Conditional Variational Autoencoder (CCVAE) to generate synthetic target-domain image features for unseen classes.
• Conducted extensive experiments on three domain adaptation datasets, including a custom X-ray security checkpoint dataset.

Main Results:

• The proposed CCVAE approach demonstrated significant effectiveness in domain adaptation tasks with partial target class availability.
• The method successfully generated synthetic features for unseen classes, improving model performance.
• Evaluations against established benchmarks confirmed the superiority of the novel approach.

Conclusions:

• The CCVAE offers a viable solution for domain adaptation problems with unseen classes in the target domain.
• The approach has practical implications, particularly in fields like aviation security requiring adaptation to new threats or object classes.
• This work advances the integration of domain adaptation and zero-shot learning for more robust and adaptable AI systems.