Adversarial Domain Adaptation With Prototype-Based Normalized Output Conditioner.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation (UDA) is crucial for aligning data across domains.
• Existing domain adversarial training methods often require complex designs or are inefficient.
• There is a need for simpler, more effective UDA techniques.

Purpose of the Study:

• To develop simple and compact conditional domain adversarial training methods for UDA.
• To improve the effectiveness of conditional alignment in domain adversarial training.
• To enhance feature alignment across domains for better adaptation performance.

Main Methods:

• Introduced Normalized OutpUt coNditioner (NOUN) by normalizing output predictions for stronger conditional alignment.
• Developed PROtotype-based Normalized OutpUt coNditioner (PRONOUN) by conditioning alignment in the prototype space.
• Utilized object recognition and semantic segmentation tasks for evaluation.

Main Results:

• NOUN effectively aligns multi-modal structures and outperforms existing UDA methods.
• PRONOUN further enhances adaptation performance by incorporating prototype-based conditioning.
• Both methods demonstrate significant improvements on UDA benchmarks.

Conclusions:

• NOUN and PRONOUN offer efficient and effective solutions for UDA.
• Prototype-based conditioning is a promising direction for improving domain adversarial training.
• The proposed methods advance the state-of-the-art in unsupervised domain adaptation.