Deep Domain Generalization via Conditional Invariant Adversarial Networks

Abstract

Domain generalization aims to apply knowledge gained from multiple labeled source domains to unseen target domains. The main difficulty comes from the dataset bias: training data and test data have different distributions, and the training set contains heterogeneous samples from different distributions. Let X denote the features, and Y be the class labels. Existing domain generalization methods address the dataset bias problem by learning a domain-invariant representation h(X) that has the same marginal distribution P(h(X)) across multiple source domains. The functional relationship encoded in P(Y |X) is usually assumed to be stable across domains such that P(Y |h(X)) is also invariant. However, it is unclear whether this assumption holds in practical problems. In this paper, we consider the general situation where both P(X) and P(Y |X) can change across all domains. We propose to learn a feature representation which has domain-invariant class conditional distributions P(h(X)|Y ). With the conditional invariant representation, the invariance of the joint distribution P(h(X), Y ) can be guaranteed if the class prior P(Y ) does not change across training and test domains. Extensive experiments on both synthetic and real data demonstrate the effectiveness of the proposed method.

Domain Generalization by Solving Jigsaw Puzzles

“…TF is the low-rank parametrized network that was presented together with the dataset PACS in [27]. CIDDG is the conditional invariant deep domain generalization method presented in [29] that trains for image classification with two adversarial constraints: one that maximizes the overall domain confusion following [19] and a second one that does the same per-class. In the DeepC variant, only this second condition is enabled.…”

Correlation-aware adversarial domain adaptation and generalization

“…For domain generalization, the training data always contains more than one source domain. Most of the existing domain generalization methods [39,20,22,38] split the source data as 70% -30%…”

“…Office-Caltech, we compare our method on the DG scenario with the state-of-the-art DG methods: learned -support vector machine (L-SVM) [47], kernel fisher discriminant analysis (KDA) [49], domain-invariant component analysis (DICA) [36], multi-task auto-encoder (MTAE) [20], domain separation network (DSN) [48], deeper, broader and artier domain generalization (DBADG) [21], conditional invariant deep domain generalization (CIDDG) [38], undoing the damage of dataset bias (Undo-Bias) [19], unbiased metric learning (UML) [46], multi-task autoencoders (MTAE) [20] and deep domain generalization with structured low-rank constraint (DGLRC) [22].…”

Single-Shot 3D Mesh Estimation via Adversarial Domain Adaptation

“…However, in practice, weight sharing reduces the number of parameters to be learned and GPU memory requirement while training. Joint optimization of deep networks on inputs from multiple domains has been studied in several recent works [21,33,41,43,73,94,102]. Genova et al [24] use unsupervised regression for 3D face modelling.…”