Synthetic Source Universal Domain Adaptation through Contrastive Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Universal Domain Adaptation (UDA) is vital for training deep learning models across diverse imaging sensors.
• Challenges in UDA include unlabeled target data and lack of prior domain knowledge, leading to model degradation.
• Model degradation in the target domain stems from insufficient discriminative power of target features.

Purpose of the Study:

• To address the degradation of models in universal domain adaptation (UDA) when using unlabeled target data.
• To propose a UDA method that enhances the discriminative power of target domain features without prior knowledge.
• To enable effective model training using synthetic source data and real target data.

Main Methods:

• A novel UDA method incorporating target domain contrastive learning is proposed.
• The method trains the discriminativeness of target features in an unsupervised manner.
• A shared feature extraction network for both source and target domains prevents computational overhead.

Main Results:

• The proposed UDA method significantly outperforms baseline methods on benchmark datasets.
• Performance improvements of 2.7% on VisDa-2017 and 5.1% on MNIST to SVHN were achieved.
• The method effectively leverages synthetic source data for training on real target data.

Conclusions:

• Target domain contrastive learning is a viable strategy to improve UDA performance.
• The proposed method offers an efficient and effective solution for domain adaptation challenges.
• This approach enhances model robustness and generalizability across different data domains.