Unsupervised Domain Adaptation via Bidirectional Transmission Generator Self-Training.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) seeks to leverage labeled source data for unlabeled target domains.
• Self-training methods are effective in UDA but can be hindered by source domain information in the feature space.
• This can trap models in the source domain, impeding the learning of target-specific features.

Purpose of the Study:

• To propose a novel self-training domain adaptation (DA) model using bidirectional transmission generators (BDTGs).
• To address the challenge of learning discriminative target domain features in UDA.
• To improve classification performance in target domains by mitigating source domain bias.

Main Methods:

• Introduced a self-training DA model with bidirectional transmission generators (BDTGs).
• Employed a bidirectional transmission structure for generators, utilizing exponential moving average (EMA) as a bridge.
• Implemented a cosine similarity-based filter to identify and mitigate the impact of noisy pseudo-labels.

Main Results:

• The bidirectional transmission structure facilitates feature space shifting, aiding target domain adaptation.
• The transmission mechanism disturbs classification boundaries, revealing unreliable target samples.
• The cosine similarity filter effectively reduces the influence of semantically incorrect pseudo-labels.
• Experiments on five benchmark UDA datasets demonstrated superior classification performance.

Conclusions:

• The proposed BDTGs effectively address the limitations of traditional self-training in UDA.
• The bidirectional transmission and filtering mechanism enhance the model's ability to adapt to target domains.
• The approach achieves state-of-the-art results on benchmark UDA datasets.