Data augmentation with norm-AE and selective pseudo-labelling for unsupervised domain adaptation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised Domain Adaptation (UDA) is crucial for applying models to new data distributions.
• Existing UDA methods often focus on data distribution alignment or learning domain-invariant features.

Purpose of the Study:

• To propose a new perspective on UDA by learning a unified classifier without explicit domain alignment.
• To leverage unlabeled target domain data effectively using Selective Pseudo-Labelling (SPL).
• To enhance classifier training through data augmentation with a novel generative model, norm-AE.

Main Methods:

• Directly learning a unified classifier in a high-dimensional feature space.
• Employing Selective Pseudo-Labelling (SPL) to utilize unlabeled target domain samples.
• Proposing norm-AE, a generative model for synthetic feature generation as data augmentation.

Main Results:

• The SPL strategy alone achieves performance comparable to state-of-the-art methods.
• The norm-AE generative model further improves performance as a data augmentation technique.
• Achieved high average accuracy on benchmark datasets like Office-Caltech (93.4%) and ImageCLEF-DA (90.4%).

Conclusions:

• The proposed methods, naive-SPL and norm-AE-SPL, offer effective UDA solutions.
• Simple classifiers trained in the original feature space can handle domain discrepancies.
• The approach provides a computationally efficient and high-performing alternative for UDA.