Heterogeneous Domain Adaptation by Information Capturing and Distribution Matching.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Heterogeneous domain adaptation (HDA) addresses challenges arising from differing feature representations between source and target domains.
• Existing HDA methods often focus on mapping matrices, potentially overlooking information loss during feature transformation.
• The reconstruction error, a measure of information loss, is often inadequately considered in conventional HDA approaches.

Purpose of the Study:

• To propose a novel HDA method that simultaneously preserves information and aligns domain distributions in a latent feature space.
• To introduce a learning model that minimizes both reconstruction loss and Maximum Mean Discrepancy (MMD).
• To address the challenge of differing feature representations in HDA.

Main Methods:

• The proposed method jointly captures information and matches source/target domain distributions in a latent feature space.
• Minimization of reconstruction loss ensures information preservation during feature transformation.
• Reduction of Maximum Mean Discrepancy aligns the distributions of the source and target domains.
• A generalized gradient flow method is employed to solve the optimization problem involving orthogonal constraints.

Main Results:

• Extensive experiments were conducted on multiple image classification datasets.
• The proposed method demonstrated superior effectiveness compared to state-of-the-art HDA techniques.
• The efficiency of the developed method was also validated against existing approaches.

Conclusions:

• The proposed method offers an effective and efficient solution for heterogeneous domain adaptation.
• Jointly minimizing reconstruction loss and MMD is a viable strategy for HDA.
• The generalized gradient flow approach successfully handles orthogonal constraints in the learning process.