Unsupervised Domain Adaptation with Asymmetrical Margin Disparity loss and Outlier Sample Extraction

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation (UDA) transfers knowledge from labeled source domains to unlabeled target domains.
• Existing UDA methods struggle with classifying similar target samples and mitigating negative transfer from source domain outliers.
• There's a need for UDA techniques that enhance feature distinction and handle domain discrepancies effectively.

Purpose of the Study:

• To propose a novel UDA method, AMD-Net with OSE, to address limitations in current approaches.
• To improve feature extraction capabilities and prevent models from converging to local optima.
• To effectively handle confusing samples in the target domain and mitigate negative transfer from source domain outliers.

Main Methods:

• Developed an Asymmetrical Margin Disparity Discrepancy (AMD) method utilizing a multi-label entropy metric for confusing target samples.
• Implemented a progressive learning scheme with a sample selection mechanism, applying different loss functions based on sample certainty.
• Introduced an outlier sample extraction (OSE) algorithm using weighted cosine similarity to reduce negative transfer from source domain outliers.

Main Results:

• AMD-Net with OSE demonstrated superior performance across four benchmark datasets: Office-31, Office-Home, VisDA-2017, and DomainNet.
• The method effectively improved feature extraction and successfully handled confusing samples and source domain outliers.
• Achieved state-of-the-art results in various unsupervised domain adaptation scenarios.

Conclusions:

• AMD-Net with OSE offers a robust solution for unsupervised domain adaptation challenges.
• The proposed AMD and OSE components significantly enhance model performance and generalization.
• This work advances the field of UDA by providing a more effective approach to domain transfer.