Learning Task Relations for Test-Time Training
Keywords: Test-time Training, Task Relation Learning, Multi-task Learning
Abstract: Generalizing deep neural networks to unseen target domains presents a major challenge in real-world deployments. Test-time training (TTT) addresses this is- sue by using an auxiliary self-supervised task to reduce the gap between source and target domains caused by distribution shifts during deployment. Previous re- search relies on the assumption that the adopted auxiliary task would be beneficial to the target task we want to adapt. However, this situation is not guaranteed as each task has a different objective, thus adaptation relies on the relation be- tween the tasks. This limitation has motivated us to introduce a more generalized framework: Task Relation Learning for Test-time Training (TR-TTT), which can be applied to multiple tasks concurrently. Our key assumption is that task re- lations are crucial information for successful test-time training, and we capture these relations using a Task Relation Learner (TRL). We model task relations as conditional probabilities by predicting the label of a target task based on the latent spaces of other task-specific features. By leveraging these relations, the network can more effectively handle distribution shifts and improve post-adaptation perfor- mance across various tasks—both classification and regression—unlike previous methods focused mainly on simple classification. To validate our approach, we ap- ply TR-TTT to conventional multi-task benchmarks, integrating it with the tradi- tional TTT experimental protocol. Our empirical results demonstrate that TR-TTT significantly outperforms state-of-the-art methods across a range of benchmarks.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3615
Loading