Joint weight optimization for partial domain adaptation via kernel statistical distance estimation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Partial Domain Adaptation (PDA) aims to transfer neural networks from a source to a target domain where the source label space includes the target label space.
• Existing methods focus on matching marginal distributions, which is suboptimal as target performance depends on joint distribution alignment.

Purpose of the Study:

• To propose a Joint Weight Optimization (JWO) approach for Partial Domain Adaptation.
• To address the limitations of marginal distribution matching by focusing on joint distribution alignment in the feature space.

Main Methods:

• Developed a Joint Weight Optimization (JWO) method to align weighted joint source and target distributions.
• Utilized L²-distance and χ²-divergence to measure joint distribution disparity.
• Proposed Kernel Statistical Distance Estimation (KSDE) to estimate these distances from data and optimize joint weights.

Main Results:

• The JWO approach effectively reduces joint distribution disparity.
• Neural network training on weighted source data using optimized weights improved performance on target domains.
• Experiments on popular datasets demonstrated the effectiveness of the proposed method.

Conclusions:

• Joint distribution alignment is crucial for effective Partial Domain Adaptation.
• The proposed JWO and KSDE methods offer a novel and effective solution for PDA.
• The approach facilitates better knowledge transfer in domain adaptation scenarios.