MiniMax Entropy Network: Learning Category-Invariant Features for Domain Adaptation

TL;DR

MMEN tackles unsupervised domain adaptation by injecting target category information into adversarial feature learning. It replaces the conventional domain classifier with an unfair category discriminator and optimizes a minimax objective over the entropy of target predictions, $H(p(y|x_t))$, to drive category-level feature alignment between domains. Target pseudo-labels guide the learning in a principled minimax game, aided by an auxiliary classifier on the source to stabilize training. Empirically, MMEN outperforms state-of-the-art baselines on ImageCLEF-DA and digits benchmarks, demonstrating effective category-wise invariance and compatibility with standard source classifiers.

Abstract

How to effectively learn from unlabeled data from the target domain is crucial for domain adaptation, as it helps reduce the large performance gap due to domain shift or distribution change. In this paper, we propose an easy-to-implement method dubbed MiniMax Entropy Networks (MMEN) based on adversarial learning. Unlike most existing approaches which employ a generator to deal with domain difference, MMEN focuses on learning the categorical information from unlabeled target samples with the help of labeled source samples. Specifically, we set an unfair multi-class classifier named categorical discriminator, which classifies source samples accurately but be confused about the categories of target samples. The generator learns a common subspace that aligns the unlabeled samples based on the target pseudo-labels. For MMEN, we also provide theoretical explanations to show that the learning of feature alignment reduces domain mismatch at the category level. Experimental results on various benchmark datasets demonstrate the effectiveness of our method over existing state-of-the-art baselines.

Figures (6)

• Figure 1: A comparison of feature alignment between previous adversarial-based method and our method. Red line and blue line denote the decision boundary for source domain and target domain respectively. Different shapes indicates different categories. Left: the features are scattered irreuglarly before adaptation, and the source samples can be classified accurately with available source labels. Middle: Previous adversarial-based methods employ domain classifiers to achieve domain confusion, thereby aligning the feature distribution globally in the level of domain. However, domain classifier contains no categorical information, therefore category shift between domains is ignored. Right: Our model utilizes an unfair category discriminator to align the target feature according to the decision boundary for the source, Threfore our model enables category-level feature alignment.
• Figure 2: An overview of our model during the learning process. Red line and blue line denote source flow and target flow respectively. The category discriminator D is an unfair multi-class classifier that classify source samles accurately while being confused about target samples.The generated target pseudo-labels $D(G(x_t))$ are directly utilized in adversarial training for feature alignment. By maximizing the category confusion on the target samples through D, the generators (G) are guided to align target features alike the source ones for high confidence. The unfairness design effectively transfers the source discriminability to the target domain. Besides, we apply classification loss to an auxiliary classifier C to enhance the categorical discriminability towards source samples
• Figure 3: The feature distribution of samples in two domains are visualized by t-SNE in the task MNIST $\rightarrow$ USPS. Source features are marked in violet, and target features marked in other different colors represent different categories. The feature representations are scattered irregularly between domains learned by the Source Only model. Classic adversarial-based method DANN only take account of global feature alignment in the level of domain. By contrast, our model consider the category-level alignment. The feature representations that belongs to the same category but in different domains are close. Hence, the decision boundary of task-specific classifier can be easily learned.
• Figure 4: Analysis of the category-level feature alignment in our model. For each category, we computer the distance of the feature center between source domain and target domain in the task MNIST $\rightarrow$ USPS. Compared with the performance learned by model pre-trained on the ImageNet, source only model and classic adversarial-based method DANN, our model successfully aligns the feature distribution in all categories.
• Figure 5: The visualization of accuracy and training loss in the task P $\rightarrow$ I and P $\rightarrow$ C on the CLEF datasets. $H$ and $L_c$ denote cross-entropy of pseudo-labels and cross-entropy with true labels respectively. $Accuracy$ denotes the performance on the classifier.