One Fits Many: Class Confusion Loss for Versatile Domain Adaptation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain Adaptation (DA) encompasses various setups like closed-set, open-set, and universal DA, each with unique label set and domain configurations.
• Existing DA methods are typically specialized for specific setups, leading to suboptimal performance when applied to others.
• This limitation hinders the development of universally applicable DA solutions.

Purpose of the Study:

• To introduce Versatile Domain Adaptation (VDA), a paradigm where a single method can address multiple DA setups without modification.
• To propose a general class confusion loss (CC-Loss) that reduces pairwise class confusion for improved transfer learning.
• To enhance the robustness of the CC-Loss by enforcing consistency under data augmentation.

Main Methods:

• Investigated the general inductive bias of class confusion in DA.
• Developed a novel Class Confusion Loss (CC-Loss) estimated from classifier predictions.
• Incorporated data augmentation to enforce confusion matrix consistency and improve invariance to distribution shifts.

Main Results:

• The proposed CC-Loss effectively reduces class confusion, leading to significant transfer gains.
• CC-Loss demonstrates competitive performance across various mainstream DA setups, including closed-set, partial-set, open-set, and universal DA.
• Experiments on 2D and 3D vision benchmarks validate the versatility and effectiveness of the CC-Loss.

Conclusions:

• The CC-Loss offers a unified approach to handle diverse Domain Adaptation challenges.
• VDA, powered by CC-Loss, represents a significant advancement over specialized DA methods.
• The method shows promise for real-world applications requiring adaptable and robust domain adaptation.