打破虚假相关：用于AU检测的不确定性驱动因果Transformer
Breaking Spurious Correlations: Uncertainty-Driven Causal Transformers for AU Detection
将Transformer注意力权重参数化为高斯分布以建模不确定性，并引入因果干预模块抑制混淆偏差，可复用于其他存在标签噪声和类别不平衡的细粒度人脸分析任务。
原文依据 We parameterize Transformer attention weights as Gaussian distributions to capture robust AU dependencies while explicitly modeling uncertainty in attention.
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论面部动作单元（AU）检测面临标注数据有限、严重的类别不平衡、标签噪声和混杂偏差等问题，这些问题往往导致过拟合和性能下降。我们提出了一个不确定性驱动因果Transformer（UDCT）框架，通过联合建模不确定性和因果干预来实现鲁棒的AU检测。具体而言，我们将Transformer注意力权重参数化为高斯分布，以捕获鲁棒的AU依赖关系，同时在注意力中显式建模不确定性。我们进一步引入了一种不确定性感知的损失重加权策略，以减轻训练过程中类别不平衡和标签噪声的影响。此外，我们加入了一个因果干预模块，以抑制依赖混杂因素的AU关联，并鼓励模型关注更稳定、偏差更少的AU关系。在BP4D和DISFA上的实验表明，UDCT在噪声、不平衡和分布偏移的设置下取得了有竞争力的性能，并具有更强的鲁棒性。
英文原文摘要 1014 字符
Facial Action Unit (AU) detection suffers from limited annotated data, severe class imbalance, label noise, and confounding biases, which often lead to overfitting and degraded performance. We propose an Uncertainty-Driven Causal Transformer (UDCT) framework for robust AU detection by jointly modeling uncertainty and causal intervention. Specifically, we parameterize Transformer attention weights as Gaussian distributions to capture robust AU dependencies while explicitly modeling uncertainty in attention. We further introduce an uncertainty-aware loss reweighting strategy to alleviate the effects of class imbalance and label noise during training. In addition, we incorporate a causal intervention module to suppress confounder-dependent AU associations and encourage the model to focus on more stable and less biased AU relationships. Experiments on BP4D and DISFA demonstrate that UDCT achieves competitive performance with stronger robustness under noisy, imbalanced, and distribution-shifted settings.