TTT++: When Does Self-Supervised Test-Time Training Fail or Thrive?

453 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Efficient Test-Time Model Adaptation without Forgetting (2022, methodology): “Recently, several attempts (Sun et al., 2020; Wang et al., 2021; Liu et al., 2021; Zhang et al., 2021b;c; Wang et al., 2022) have been proposed to handle the distribution shifts by online adapting a model at test time (called test-time adaptation ).”
A Comprehensive Survey on Test-Time Adaptation Under Distribution Shifts (2023, unknown): “Some approaches [272], [273] achieve feature alignment using different orders of moments for features in the source domain.”
NOTE: Robust Continual Test-time Adaptation Against Temporal Correlation (2022, background): “There also exist works [35, 24] utilizing domain-specific self-supervision to resolve the distribution shift with test data, but are complementary to ours, i.”
MEMO: Test Time Robustness via Adaptation and Augmentation (2021, background): “This failure mode may potentially be rectified by carefully choosing which parameters to adapt, such as only adapting the parameters in BN layers [46], or regularizing the model such that it does not change too drastically from the pretrained model [26].”
Towards Stable Test-Time Adaptation in Dynamic Wild World (2023, methodology): “, Test-Time Training (TTT) (Sun et al., 2020; Liu et al., 2021) and Fully TTA (Wang et al.”