Heterogeneous Domain Adaptation Through Progressive Alignment.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Real-world transfer learning, particularly cross-modal applications, faces challenges from heterogeneous domain adaptation (HDA) where source and target domains differ in features and distributions.
• Existing HDA methods often optimize either feature discrepancy or distribution divergence, potentially reinforcing the other issue.

Purpose of the Study:

• To propose a novel HDA method that unifies the optimization of both feature discrepancy and distribution divergence within a single objective function.
• To develop a method capable of handling diverse features with arbitrary dimensions, overcoming limitations of previous HDA approaches.

Main Methods:

• Introduced progressive alignment, a novel HDA technique.
• Employed dictionary-sharing coding to learn a new transferable feature space.
• Aligned distribution gaps within this newly learned feature space.

Main Results:

• Demonstrated the ability to optimize both feature discrepancy and distribution divergence concurrently.
• Showcased adaptability to diverse features across arbitrary dimensions.
• Achieved superior performance compared to state-of-the-art methods in various transfer learning tasks.

Conclusions:

• The proposed progressive alignment method offers a unified and effective solution for heterogeneous domain adaptation.
• This approach significantly advances cross-modal learning by successfully bridging domain gaps in image classification, text categorization, and text-to-image recognition.