Due to the domain shift, machine learning systems typically fail to generalize well to domains different from those of training data, which is the problem that domain generalization (DG) aims to address. However, most mainstream DG algorithms lack interpretability and require domain labels, which are not available in many real-world scenarios. In this work, we propose a novel DG method, HMOE: Hypernetwork-based Mixture of Experts (MoE), that does not require domain labels and is more interpretable. We use hypernetworks to generate the weights of experts, allowing experts to share some useful meta-knowledge. MoE has proven adept at detecting and identifying heterogeneous patterns in data. For DG, heterogeneity exactly arises from the domain shift. We compare HMOE with other DG algorithms under a fair and unified benchmark-DomainBed. Extensive experiments show that HMOE can perform latent domain discovery from data of mixed domains and divide it into distinct clusters that are surprisingly more consistent with human intuition than original domain labels. Compared to other DG methods, HMOE shows competitive performance and achieves SOTA results in some cases without using domain labels.
翻译:由于域变, 机器学习系统通常无法向与培训数据不同的领域推广, 而培训数据则是领域通用( DG) 想要解决的问题。 然而, 大多数主流的DG算法缺乏可解释性, 需要域变标签, 在许多现实世界情景中并不存在。 在这项工作中, 我们提出了一个新型的DG方法, HMOE: 超网络专家混合体(MOE), 不需要域名标签, 并且更容易解释。 我们使用超网络来生成专家的权重, 让专家分享一些有用的元知识。 与其它DG方法相比, HMOE 已经证明能够探测和识别数据中的差异模式。 对于DG, 域变异性完全产生于域变换。 我们比较HMOE 和其他DG算法, 在一个公平和统一的基准- DomainBed 下。 广泛的实验显示, HMOE 可以在混合域数据中进行隐性域发现, 并将其分为不同组群, 令人惊讶地更符合人类直觉而非原始域标。 与其他 DG 方法相比, HMOE 显示竞争性的性业绩, 并在某些情况下在没有域内取得SOTA结果 。