面向开集识别的稳健测试时适应方法

开集识别旨在研究测试阶段突现未见类别对于机器学习模型的挑战,以期学习模型既能分类已见类别又可识别/拒绝未见类别,是确保机器学习模型能够在开放世界中高效稳健部署的重要技术.既有开集识别技术通常假设已见类别的协变量分布在训练与测试阶段维持不变.然而在实际场景中,类别的协变量分布常不断变化.直接利用既有技术不再奏效,其性能甚至劣于基线方案.因此,亟需研究新型开集识别方法,使其能不断适应协变量分布偏移,以期模型在测试阶段既能稳健分类已见类别又可识别未见类别.将此新问题设置命名为开放世界适应问题(AOW),并提出了一种开放测试时适应方法(OTA).该方法基于无标注测试数据优化自适应熵损失与开集熵损失更新模型,维持对已见类的既有判别能力,同时增强了识别未见类的能力.大量实验分析表明,该方法在多组基准数据集、多组不同协变量偏移程度下均稳健地优于现有先进的开集识别方法.

Towards Robust Test-Time Adaptation for Open-Set Recognition

Open-set recognition is an important issue for ensuring the efficient and robust deployment of machine learning models in the open world. It aims to address the challenge of encountering samples from unseen classes that emerge during testing, i.e., to accurately classify the seen classes while identifying and rejecting the unseen ones. Current open-set recognition studies assume that the covariate distribution of the seen classes remains constant during both training and testing. However, in practical scenarios, the covariate distribution is constantly shifting, which can cause previous methods to fail, and their performance may even be worse than the baseline method. Therefore, it is urgent to study novel open-set recognition methods that can adapt to the constantly changing covariate distribution so that they can robustly classify seen categories and identify unseen categories during testing. We name this novel problem Adaptation in the Open World (AOW) and propose a test-time adaptation method for open-set recognition called Open-set Test-time Adaptation (OTA). OTA method only utilizes unlabeled test data to update the model with adaptive entropy loss and open-set entropy loss, maintaining the model''s ability to discriminate seen classes while further enhancing its ability to recognize unseen classes. We conduct comprehensive experiments on multiple benchmark datasets with different covariate shift levels. The results show that our proposal is robust to covariate shift and gives superior performance compared to many state-of-the-art methods.