Artificial intelligence
Test-time adaptation
Definition
Test-time adaptation adjusts a trained model using data encountered during evaluation or deployment. Unlike ordinary fixed-model inference, it updates model parameters or statistics in response to the target data.
Also known as: Test time adaptation, TTA
Updated
Adapt after the original training stage
Tent studies fully test-time adaptation with access only to test data and the trained model's parameters. It estimates normalization statistics and updates channel-wise affine parameters by minimizing prediction entropy.
Entropy here measures uncertainty in the model's output distribution. The method uses that quantity as an optimization signal without requiring target class labels during adaptation.
The setting differs from ordinary fine-tuning
Conventional domain adaptation may allow source data, labeled target examples, or a separate adaptation training stage. A fully test-time method has a more restricted information setting.
It also differs from spending more computation on a fixed model's answer. The defining feature is adaptation of the model or its statistics, not simply additional sampling or a longer reasoning process.
Perception benchmarks do not establish robot safety
Tent reports image-classification and semantic-segmentation evaluations. A robot camera encountering changed lighting is a possible application context, but the paper does not establish safe autonomous robot control under arbitrary changes.
Lower prediction entropy means greater confidence, not a direct physical correctness measurement. An implementation should identify what is updated and evaluate behavior under its intended sequence of conditions, because the deployed model changes as it encounters data.
Sources
Related terms
Domain adaptation
Domain adaptation adjusts a learned model to work on a target data distribution that differs from its source training distribution. Robotics examples include adapting perception from simulation to camera images or adapting behavior to changed physical conditions.
Self-supervised learning
Self-supervised learning builds a training signal from the structure of the data itself rather than requiring a human label for every example. In robotics it can learn useful visual or temporal representations before a downstream task policy is trained.
Catastrophic forgetting
Catastrophic forgetting is a substantial loss of previously learned capability when a model is trained on new tasks or data. It is a central problem in sequential and continual learning.