基于鲁棒和可靠对称交叉熵的测试时适应算法

测试时间适应（test-time adaptation，TTA）的目标是利用未标记的测试数据使已训练完成的神经网络模型在测试时适应测试数据分布。现有的TTA方法主要考虑在单个或多个静态环境中进行适应。然而，在非平稳环境中，测试数据分布会随着时间的推移而连续变化，这导致以往的TTA方法不稳定。因此，提出了一种基于鲁棒和可靠对称交叉熵的测试时适应算法（robust and reliable symmetric cross entropy test-time adaptation，RRSTA）。首先，为提高对噪声分布变化的鲁棒性和缓解灾难性遗忘，提出了基于均值教师模型的对称交叉熵，既鼓励模型正确预测又惩罚错误的预测。其次，为了提高对不同噪声样本的鲁棒性，提出了一种双流扰动技术，通过教师模型强视图，指导学生模型的由弱到强的扰动视图。最后，提出了可靠熵最小化策略，防止参数的剧烈变化，以稳定适应。广泛的实验和消融研究在CIFAR10C和CIFAR100C上证实了所提出方法的有效性，相比于未经适应的模型，错误率降低了26.13%和14.69%，并且显著优于次优的方法。

Robust and reliable symmetric cross-entropy-based test-time adaptation

Test-time adaptation (TTA) aims to make the trained neural network model adapt to the test data distribution at test time using unlabeled test data. Existing TTA methods mainly consider adaptation in a single or multiple static environments. However, in non-stationary environments, the test data distribution changes continuously over time, which leads to the instability of previous TTA methods. Therefore, this paper proposed a test-time adaptation algorithm(RRSTA) based on robust and reliable symmetric cross entropy. First, in order to improve the robustness to noise distribution changes and alleviate catastrophic forgetting, it proposed a symmetric cross entropy based on the mean teacher model, which encouraged the model to predict correctly and punished the wrong prediction. Secondly, in order to improve the robustness to different noise samples, it proposd a dual-stream perturbation technique, which guided the weak-to-strong perturbation view of the student model through the strong view of the teacher model. Finally, it proposed a reliable entropy minimization strategy to prevent the drastic change of parameters and stabilize adaptation. Extensive experiments and ablation studies on CIFAR10C and CIFAR100C confirm the effectiveness of the proposed method. Compared with the unadapted models, the error rate is significantly reduced by 26.13% and 14.69%, and it is significantly better than the second-best method.