Dynamic Instance Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) typically assumes coarse-grained domains with domain labels.
• Existing UDA methods struggle with fine-grained domains where distributions differ significantly.
• Coarse-grained domain alignment can be ineffective for complex, nuanced datasets.

Purpose of the Study:

• To propose a novel approach for unsupervised domain adaptation that addresses the limitations of coarse-grained domain assumptions.
• To introduce dynamic instance domain adaptation (DIDA) by treating each instance as a distinct domain.
• To develop a model that performs effective adaptation without relying on domain annotations.

Main Methods:

• Developed a dynamic neural network with adaptive convolutional kernels for instance-adaptive residuals.
• Proposed dynamic instance domain adaptation (DIDA) to adapt domain-agnostic features to individual instances.
• Employed a semi-supervised learning paradigm with cross-entropy loss on source and pseudo-labeled target data.

Main Results:

• Achieved state-of-the-art performance on multiple single-source and multi-source UDA benchmarks.
• Demonstrated the effectiveness of instance-level adaptation over traditional domain alignment.
• Showcased the ability to apply a shared classifier across source and target domains without domain labels.

Conclusions:

• Dynamic instance domain adaptation (DIDA) offers a more effective strategy for UDA, especially with fine-grained domains.
• DIDA-Net provides a robust and efficient solution for unsupervised domain adaptation challenges.
• The proposed method advances the field by eliminating the need for domain annotations and complex alignment losses.