Center transfer for supervised domain adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain adaptation (DA) is crucial for pattern recognition, leveraging source data for target domain tasks.
• Supervised domain adaptation (SDA) is valuable when target domain labeled data is scarce and expensive to collect.
• Existing SDA methods often require paired training samples, limiting their applicability.

Purpose of the Study:

• To propose a novel supervision signal, center transfer loss (CTL), for efficient feature alignment in deep learning-based SDA.
• To enhance the performance of deep learning models in target domains with limited labeled data.
• To address limitations of existing SDA methods by removing the need for paired samples and balancing hyper-parameters.

Main Methods:

• Developed a new supervision signal: center transfer loss (CTL).
• Implemented CTL using a one-stream input mini-batch strategy, eliminating the need for sample pairing.
• CTL integrates domain alignment and feature discriminative power enhancement within the training process.

Main Results:

• CTL demonstrated improved performance in deep learning models under SDA settings.
• The proposed method achieved superior results compared to recent state-of-the-art approaches on public datasets.
• CTL effectively aligns features and increases their discriminative power without requiring a balancing hyper-parameter.

Conclusions:

• Center transfer loss (CTL) offers an efficient and effective approach for supervised domain adaptation in deep learning.
• CTL provides a simplified yet powerful method for improving model generalization in data-scarce target domains.
• The proposed method represents a significant advancement over existing SDA techniques.