[Paper Review] Opposite Structure Learning for Semi-supervised Domain Adaptation

Can Qin, Lichen Wang+4 more|arXiv (Cornell University)|Feb 6, 2020

Domain Adaptation and Few-Shot Learning32 references20 citations

TL;DR

This paper proposes Unifying Opposite Structures for Domain Adaptation (UODA), a semi-supervised method that alternates between expanding source features and clustering target features to align cross-domain distributions. By training source-based and target-based classifiers with opposing loss forms, UODA improves decision boundary smoothness and inter-class separation, achieving state-of-the-art performance on DomainNet and Office-Home.

ABSTRACT

Current adversarial adaptation methods attempt to align the cross-domain features whereas two challenges remain unsolved: 1) conditional distribution mismatch between different domains and 2) the bias of decision boundary towards the source domain. To solve these challenges, we propose a novel framework for semi-supervised domain adaptation by unifying the learning of opposite structures (UODA). UODA consists of a generator and two classifiers (i.e., the source-based and the target-based classifiers respectively) which are trained with opposite forms of losses for a unified object. The target-based classifier attempts to cluster the target features to improve intra-class density and enlarge inter-class divergence. Meanwhile, the source-based classifier is designed to scatter the source features to enhance the smoothness of decision boundary. Through the alternation of source-feature expansion and target-feature clustering procedures, the target features are well-enclosed within the dilated boundary of the corresponding source features. This strategy effectively makes the cross-domain features precisely aligned. To overcome the model collapse through training, we progressively update the measurement of distance and the feature representation on both domains via an adversarial training paradigm. Extensive experiments on the benchmarks of DomainNet and Office-home datasets demonstrate the effectiveness of our approach over the state-of-the-art method.

Motivation & Objective

• Address the conditional distribution mismatch between source and target domains in semi-supervised domain adaptation.
• Reduce the bias of the decision boundary toward the source domain by improving target feature clustering.
• Enhance feature alignment by alternately expanding source features and clustering target features.
• Prevent model collapse during adversarial training through progressive updates to distance measurement and feature representation.
• Achieve precise cross-domain feature alignment without relying on adversarial domain alignment alone.

Proposed method

• Introduce a generator and two classifiers: a source-based classifier that scatters source features to improve decision boundary smoothness.
• Design a target-based classifier that clusters target features to increase intra-class density and inter-class divergence.
• Train the two classifiers with opposing loss functions to unify the learning of opposite structures in a single optimization objective.
• Alternate between source-feature expansion and target-feature clustering procedures to dilate the source boundary and enclose target features within it.
• Apply adversarial training to progressively update feature representations and distance measurements, mitigating model collapse.
• Use a unified objective that jointly optimizes feature alignment and decision boundary robustness across domains.

Experimental results

Research questions

• RQ1 How can we effectively reduce conditional distribution mismatch between source and target domains in semi-supervised domain adaptation?
• RQ2 To what extent can opposing loss forms for source and target classifiers improve decision boundary generalization?
• RQ3 Can alternating source expansion and target clustering procedures lead to better cross-domain feature alignment than standard adversarial alignment?
• RQ4 How does progressive refinement of feature representation and distance measurement prevent model collapse during training?
• RQ5 Does the proposed UODA framework outperform existing state-of-the-art methods on standard domain adaptation benchmarks?

Key findings

• UODA achieves state-of-the-art performance on the DomainNet benchmark, outperforming existing methods in cross-domain generalization.
• The method significantly improves target feature clustering, leading to higher intra-class compactness and inter-class separability.
• By opposing the loss forms on source and target classifiers, UODA reduces decision boundary bias toward the source domain.
• The alternation between source expansion and target clustering results in precise alignment of cross-domain features.
• Progressive updates to distance measurement and feature representation effectively prevent model collapse during adversarial training.
• Extensive experiments on Office-Home confirm that UODA generalizes well across diverse domain shifts with minimal supervision.