IEL: Intra-Model Ensemble Learning For Single Sample Test-Time Adaptation
Keywords: Test-Time Adaptation, Ensemble Learning, Entropy-Regularization, Knowledge Distillation
TL;DR: Ensemble and Entropy-Optimization based algorithm for adapting sets of pre-trained models to distribution shifted data.
Abstract: Test-Time Adaptation (TTA) problems involve adapting pre-trained models to new data distributions in testing time, with access to only model weights and a stream of unlabeled data. In this work, we present IEL, a method for adapting sets of independently pre-trained classifiers to distribution shifted data one sample at a time without labels. We minimize the cross-entropy between the classifier output that has the highest predicted probability for the majority voted class (a high confidence softmax) and all other models in a set of classifiers. The majority voted model that all others learn from may change from sample to sample, allowing the group to collectively learn from each other. Our method uniquely optimizes all trainable parameters in each model and needs only a single sample for adaptation. Using sets of independently pre-trained base classifiers with distinct architectures, we show that our approach can reduce generalization error for image classification tasks on corrupted CIFAR-10, CIFAR-100, and ImageNet while also minimizing the entropy of model outputs.
Primary Area: transfer learning, meta learning, and lifelong 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.
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: 5005
Loading