Learning Domain-Independent Deep Representations by Mutual Information Minimization.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Domain transfer learning seeks common representations for source and target domains.
• Conventional methods match domain distributions, relying on accurate characterization and matching criteria.

Purpose of the Study:

• To propose a novel framework for domain transfer representation learning.
• To develop domain-independent data representations for enhanced cross-domain generalization.

Main Methods:

• Minimize mutual information between representations and domain indicators.
• Utilize a classwise deep convolutional network to maximize class margins.
• Employ an iterative algorithm based on Adam optimization for simultaneous parameter and representation learning.

Main Results:

• Learned representations are independent of their original domains.
• Achieved effective domain transfer learning with improved classification.
• Demonstrated advantage over existing domain transfer learning methods through extensive experiments.

Conclusions:

• The proposed method learns discriminate and domain-independent representations.
• This approach enhances the generalizability of learned features across different domains.
• The framework offers a promising direction for advancing domain transfer learning techniques.