A Co-Training Framework for Heterogeneous Heuristic Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation (UDA) methods often rely on fixed neural network structures.
• Existing UDA techniques may not adequately address class discriminative representations.
• Domain-specific properties can hinder effective model generalization.

Purpose of the Study:

• To propose a co-training framework for heterogeneous heuristic domain adaptation (CO-HHDA).
• To enhance class discriminative representations in UDA.
• To address limitations in current UDA methods regarding domain-specific information and class separability.

Main Methods:

• Introduced a heterogeneous heuristic network to model domain-specific characteristics, allowing flexible network structures per domain.
• Developed a co-training scheme with a source classifier and a target classifier trained on domain-invariant representations.
• Utilized adaptive thresholds for selecting reliable pseudolabels to facilitate mutual learning between classifiers.

Main Results:

• Empirical results on three benchmark datasets show CO-HHDA outperforms state-of-the-art domain adaptation methods.
• The proposed framework effectively learns domain-invariant representations while preserving class discriminative information.
• Heterogeneous heuristic networks adapt better to domain-specific information without overfitting or underfitting.

Conclusions:

• The CO-HHDA framework offers a significant advancement in unsupervised domain adaptation.
• Mutual learning between classifiers and adaptive pseudolabel selection are key to improving representation learning.
• This approach provides a more robust and effective solution for heterogeneous domain adaptation tasks.