期刊
KNOWLEDGE-BASED SYSTEMS
卷 283, 期 -, 页码 -出版社
ELSEVIER
DOI: 10.1016/j.knosys.2023.111164
关键词
Deep learning; Test time adaptation; Distribution shift; Robustness
In this paper, a novel method called CSTTA is proposed for test time adaptation (TTA), which utilizes confidence-based optimization and sample reweighting to better utilize sample information. Extensive experiments demonstrate the effectiveness of the proposed method.
Recently, Test Time Adaptation (TTA) has emerged as a solution to real-world challenges posed by inconsistencies between training and testing distributions. This technique refines a pretrained model for a target domain using only unlabeled test data. However, most existing TTA methods assign equal weights to all samples and are dominated by entropy minimization for adaptation. In this paper, we propose a novel confidence-based optimization strategy and theoretically show that the proposed method tends to yield larger gradients than entropy-based methods and has the potential to enhance performance. Moreover, we show that the importance of samples is frequently underestimated and propose a novel truncation function that assigns an adaptive weight to each sample. The proposed method, named CSTTA, consists of a novel confidence-based optimization strategy and sample-reweighted strategy. It aims to better utilize the sample information for quicker adaptation to new scenarios. Extensive experiments on three digital datasets (CIFAR10-C, CIFAR100-C and ImageNet-C) and a real-world dataset (ImageNet-3DCC) demonstrate the effectiveness of the proposed method.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据