Prototypical Influence Function for Fully Test-time Adaptation
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.
Keywords: test-time adaptation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Test-time adaptation (TTA) addresses domain shift issues in real-world applications. TTA adapts the model considering real-world constraints: (1) TTA does not have access to the training data or the labels of the test data and (2) TTA has limited computational resources for adaptation since it adapts model while performing inference. Due to the constraints, it has been established that model updates based on model-trusting data whose predictions closely aligned with one-hot vectors are effective. Hence, we propose a PIF regularizer utilizing the influence function to assess the influence of adapting a test data point on the loss for model-trusting data. The influence function is impractical for TTA due to computational complexity and the unavailability of model-trusting data. However, by introducing reasonable
approximations, we can feasibly use the PIF for TTA. Our experimental results demonstrate consistent performance enhancement when the PIF is applied into the existing TTA methods on various benchmark datasets.
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: 1840
Loading