Open-Set Hypothesis Transfer With Semantic Consistency.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised open-set domain adaptation (UODA) addresses scenarios with unlabeled target data containing unknown classes.
• Existing UODA methods require source domain data, limiting applications due to privacy concerns.
• The hypothesis transfer setting for UODA, where source data is unavailable, presents a significant challenge.

Purpose of the Study:

• To develop a novel UODA method for the hypothesis transfer setting, overcoming the absence of source domain data.
• To enable accurate classification of unlabeled target data into known and unknown classes.

Main Methods:

• The proposed method utilizes pseudo-labels generated from confident predictions on target data.
• A novel consistency regularization technique is applied to semantically transformed inputs to discover latent class semantics.
• The model enforces consistent and definite predictions on transformed data.

Main Results:

• The method successfully classifies unlabeled data into discriminative classes, including unknown categories.
• Theoretical analysis proves that the consistency objective can recover true label information under ideal semantic transformations.
• Experimental results demonstrate superior performance compared to state-of-the-art UODA methods on benchmark datasets.

Conclusions:

• The proposed UODA approach effectively handles the hypothesis transfer setting without source data.
• Pseudo-labeling and consistency regularization are key components for successful adaptation in open-set scenarios.
• This work advances UODA by enabling adaptation with restricted source data access.