[Paper Review] Domain Adaptation and Image Classification via Deep Conditional Adaptation Network

Pengfei Ge, Chuan-Xian Ren+2 more|arXiv (Cornell University)|Jun 14, 2020

Domain Adaptation and Few-Shot Learning46 references4 citations

TL;DR

This paper proposes Deep Conditional Adaptation Network (DCAN), a novel unsupervised domain adaptation method that aligns conditional distributions of deep features between source and target domains using Conditional Maximum Mean Discrepancy (CMMD) and enhances discriminative feature learning via mutual information maximization. DCAN achieves state-of-the-art performance on both standard and partial domain adaptation benchmarks, outperforming existing methods by mitigating negative transfer and improving robustness to label distribution shifts.

ABSTRACT

Unsupervised domain adaptation aims to generalize the supervised model trained on a source domain to an unlabeled target domain. Marginal distribution alignment of feature spaces is widely used to reduce the domain discrepancy between the source and target domains. However, it assumes that the source and target domains share the same label distribution, which limits their application scope. In this paper, we consider a more general application scenario where the label distributions of the source and target domains are not the same. In this scenario, marginal distribution alignment-based methods will be vulnerable to negative transfer. To address this issue, we propose a novel unsupervised domain adaptation method, Deep Conditional Adaptation Network (DCAN), based on conditional distribution alignment of feature spaces. To be specific, we reduce the domain discrepancy by minimizing the Conditional Maximum Mean Discrepancy between the conditional distributions of deep features on the source and target domains, and extract the discriminant information from target domain by maximizing the mutual information between samples and the prediction labels. In addition, DCAN can be used to address a special scenario, Partial unsupervised domain adaptation, where the target domain category is a subset of the source domain category. Experiments on both unsupervised domain adaptation and Partial unsupervised domain adaptation show that DCAN achieves superior classification performance over state-of-the-art methods.

Motivation & Objective

• To address the limitation of marginal distribution alignment in unsupervised domain adaptation, which ignores conditional dependence between features and labels and risks negative transfer.
• To develop a method that effectively aligns conditional distributions $P^s(Z|Y)$ and $P^t(Z|Y)$ to reduce domain discrepancy while preserving discriminative information.
• To handle the challenging Partial UDA scenario where the target domain contains a subset of source domain categories.
• To improve feature discrimination in unlabeled target domains by leveraging mutual information between samples and predicted labels.
• To provide a robust, efficient, and scalable method for conditional domain adaptation that generalizes across diverse domain shift scenarios.

Proposed method

• DCAN minimizes Conditional Maximum Mean Discrepancy (CMMD) between the conditional distributions of deep features in source and target domains to align $P^s(Z|Y)$ and $P^t(Z|Y)$.
• The method uses a kernel-based estimation of CMMD to measure the discrepancy between conditional distributions without requiring per-class statistics or adversarial training.
• An additional mutual information maximization objective is introduced to extract discriminative information from unlabeled target samples, enhancing pseudo-label quality.
• The model is trained end-to-end using a joint loss combining CMMD minimization and mutual information maximization, enabling effective feature learning under domain shift.
• DCAN is designed to be robust to small batch sizes and does not require fine-tuning of hyperparameters like batch size.
• The framework supports both standard UDA and Partial UDA, where target categories are a subset of source categories, by avoiding misclassification of non-existent classes.

Experimental results

Research questions

• RQ1 Can conditional distribution alignment via CMMD reduce domain discrepancy more effectively than marginal distribution alignment in unsupervised domain adaptation?
• RQ2 How does incorporating mutual information between features and predictions improve model generalization on unlabeled target data?
• RQ3 Can DCAN effectively handle the Partial UDA setting where target categories are a subset of source categories, avoiding misclassification of non-existent classes?
• RQ4 How does DCAN perform compared to state-of-the-art methods like DANN and DDC in terms of accuracy and robustness across diverse domain shift scenarios?
• RQ5 What is the computational cost of CMMD estimation, and how does it scale with batch size and feature dimensionality?

Key findings

• On the Office-31 dataset, DCAN achieves classification accuracy close to 100% when using ground truth labels for CMMD estimation, significantly outperforming the pseudo-label version.
• On the more challenging W→A and D→A tasks, DCAN with ground truth labels improves accuracy by approximately 20% over the pseudo-label variant.
• In the Partial UDA setting on Office-Home (C→P), DCAN maintains over 75% accuracy even when the target domain has only 5 classes, while DANN degrades below the Source-Only baseline.
• DCAN’s training time is about 6% longer than DANN and 11% longer than DDC, primarily due to kernel matrix computation, but remains practical for real-world deployment.
• The method demonstrates robustness to small batch sizes (e.g., n=8), indicating minimal need for hyperparameter tuning in practice.
• The ablation study confirms that CMMD-based conditional alignment is more effective than marginal alignment in reducing negative transfer, especially in Partial UDA.