Revisiting Realistic Test-Time Training: Sequential Inference and Adaptation by Anchored Clustering Regularized...

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Model deployment often faces distribution shift, necessitating adaptation.
• Test-time training (TTT) addresses adaptation when source data is unavailable and rapid inference is crucial.
• Existing TTT research lacks standardized experimental settings, hindering fair comparisons.

Purpose of the Study:

• To establish clear TTT protocols and assumptions.
• To develop a robust TTT method for improved feature learning and adaptation.
• To provide a fair benchmarking framework for TTT methods.

Main Methods:

• Categorization of TTT protocols based on data streaming and source model retraining.
• Development of test-time anchored clustering (TTAC) for feature learning and domain adaptation.
• Introduction of TTAC++, a method regularizing self-training with anchored clustering for enhanced TTT.

Main Results:

• TTAC++ demonstrates superior performance across five diverse TTT datasets.
• The proposed TTAC method effectively learns features and improves adaptation, even in source-free scenarios.
• Consistent outperformance of TTAC++ over state-of-the-art methods under various TTT protocols.

Conclusions:

• Standardized TTT protocols are essential for reliable research and development.
• TTAC++ offers a significant advancement in test-time adaptation, addressing limitations of prior methods.
• This work provides a foundation for fair benchmarking and future advancements in TTT.