Unsupervised domain adaptation with deep network based on discriminative class-wise MMD

Abstract

General learning algorithms trained on a specific dataset often have difficulty generalizing effectively across different domains. In traditional pattern recognition, a classifier is typically trained on one dataset and then tested on another, assuming both datasets follow the same distribution. This assumption poses difficulty for the solution to be applied in real-world scenarios. The challenge of making a robust generalization from data originated from diverse sources is called the domain adaptation problem. Many studies have suggested solutions for mapping samples from two domains into a shared feature space and aligning their distributions. To achieve distribution alignment, minimizing the maximum mean discrepancy (MMD) between the feature distributions of the two domains has been proven effective. However, this alignment of features between two domains ignores the essential class-wise alignment, which is crucial for adaptation. To address the issue, this study introduced a discriminative, class-wise deep kernel-based MMD technique for unsupervised domain adaptation. Experimental findings demonstrated that the proposed approach not only aligns the data distribution of each class in both source and target domains, but it also enhances the adaptation outcomes.

Keywords

• maximum mean discrepancy (mmd)
• unsupervised domain adaptation
• transfer learning
• reproduced kernel hilbert space
• pseudo labels