Toward Open-World Domain Adaptation via Iteratively Contrastive Learning and Clustering

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Open-set domain adaptation (DA) addresses covariate and category shifts between source and target domains.
• Existing methods often fail to discover novel classes in the target domain, labeling them as 'unknown'.

Purpose of the Study:

• To introduce a more challenging open-world DA problem: recognizing seen classes while discovering novel classes.
• To propose a novel framework for open-world DA that leverages clustering and contrastive learning.

Main Methods:

• The framework converts the problem into a clustering task using contrastive learning to model instance relationships.
• It employs an iterative process involving semi-supervised clustering and contrastive learning steps.
• The method is optimizable via an expectation-maximization (EM) algorithm.

Main Results:

• The proposed method achieves superior performance across five public datasets.
• It effectively clusters unlabeled target data into seen and novel classes.
• Contrastive losses reduce domain discrepancy and facilitate novel class discovery.

Conclusions:

• The developed framework successfully addresses the open-world DA problem.
• This work establishes a new benchmark for future research in open-world domain adaptation.