[Paper Review] Contrastive Adaptation Network for Unsupervised Domain Adaptation

Guoliang Kang, Lu Jiang+2 more|arXiv (Cornell University)|Jan 4, 2019

Domain Adaptation and Few-Shot Learning39 references102 citations

TL;DR

CAN introduces a class-aware contrastive discrepancy (CDD) for unsupervised domain adaptation and shows state-of-the-art results on Office-31 and competitive results on VisDA-2017 by learning more discriminative, intra-class compact and inter-class separable features.

ABSTRACT

Unsupervised Domain Adaptation (UDA) makes predictions for the target domain data while manual annotations are only available in the source domain. Previous methods minimize the domain discrepancy neglecting the class information, which may lead to misalignment and poor generalization performance. To address this issue, this paper proposes Contrastive Adaptation Network (CAN) optimizing a new metric which explicitly models the intra-class domain discrepancy and the inter-class domain discrepancy. We design an alternating update strategy for training CAN in an end-to-end manner. Experiments on two real-world benchmarks Office-31 and VisDA-2017 demonstrate that CAN performs favorably against the state-of-the-art methods and produces more discriminative features.

Motivation & Objective

• Motivate unsupervised domain adaptation to leverage class information rather than only domain alignment.
• Propose a new Contrastive Domain Discrepancy (CDD) that minimizes intra-class domain discrepancy while maximizing inter-class domain discrepancy.
• Develop an end-to-end training framework (CAN) that alternates target label estimation and feature adaptation.
• Demonstrate that class-aware alignment yields more discriminative features and better generalization on benchmark datasets.

Proposed method

• Define the Contrastive Domain Discrepancy (CDD) to measure intra-class and inter-class domain differences.
• Minimize intra-class discrepancy and maximize inter-class discrepancy using an MMD-based formulation.
• Embed CDD into a deep network as a loss term over task-specific FC layers with weight β (Eq. 8).
• Train CAN via alternating optimization: cluster-target-labels via spherical K-means and then back-propagate to adapt features using CDD (Eq. 6, 3.2).
• Use class-aware sampling (CAS) to ensure mini-batches contain multiple domains for each class, enabling reliable CDD estimation.
• Adopt progressive learning by discarding ambiguous target data/classes during clustering and gradually including more classes.

Experimental results

Research questions

• RQ1 Does incorporating class information through Contrastive Domain Discrepancy improve unsupervised domain adaptation compared to class-agnostic approaches?
• RQ2 Can CAN achieve superior performance on standard UDA benchmarks (Office-31, VisDA-2017) relative to state-of-the-art methods?
• RQ3 How do intra-class compactness and inter-class separability contribute to target-domain discriminability and generalization?
• RQ4 What is the role of alternative optimization and class-aware sampling in the effectiveness of CAN?

Key findings

• CAN achieves 90.6% average accuracy on Office-31 with ResNet-50 (CAN) vs. 89.5% for intra-class only and higher than competing methods.
• On VisDA-2017 (validation set) with ResNet-101, CAN attains 87.2% mean accuracy, outperforming several baselines and reaching competitive results on the test set (87.4% with a single model).
• Class-aware alignment (CDD) improves performance over purely intra-class alignment, indicating that maximizing inter-class domain discrepancy helps generalization.
• Ablation studies show that both alternative optimization (AO) and class-aware sampling (CAS) contribute to performance gains, with CAN surpassing variants lacking these components.
• The visualization (t-SNE) suggests CAN yields higher intra-class compactness and larger inter-class margins than JAN.
• CAN achieves the best-published result on Office-31 and is very competitive on VisDA-2017, illustrating effective class-aware domain adaptation.