Multi-source fully test-time adaptation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) excel in many fields but struggle with generalization when test data distribution differs from training data.
• Existing fully test-time adaptation methods use a single trained model, limiting the potential information for adaptation.
• Real-world scenarios often involve multiple trained models that can offer complementary insights for adaptation.

Purpose of the Study:

• To address the limitations of single-model adaptation by proposing the problem of multi-source fully test-time adaptation.
• To develop a method that leverages multiple pre-trained models for enhanced adaptation to unlabeled test data.
• To improve the generalization capabilities of deep learning models in domain-shifted scenarios.

Main Methods:

• Introduced the novel problem of multi-source fully test-time adaptation.
• Developed a method employing a weighted aggregation scheme to prioritize relevant models.
• Incorporated two unsupervised losses for joint adaptation of multiple models using online unlabeled samples.

Main Results:

• The proposed method demonstrated superior performance compared to baseline approaches on three image classification datasets.
• Weighted aggregation adaptively assigned higher relevance to more suitable models.
• Joint adaptation using unsupervised losses effectively utilized complementary information from multiple models.

Conclusions:

• Multi-source fully test-time adaptation offers a promising direction for improving model generalization.
• The proposed weighted aggregation and unsupervised loss method effectively adapts multiple models.
• This approach enhances robustness and performance in scenarios with varying data distributions.