[Paper Review] A DIRT-T Approach to Unsupervised Domain Adaptation

Rui Shu, Hung Bui+2 more|arXiv (Cornell University)|Feb 23, 2018

Domain Adaptation and Few-Shot Learning36 references293 citations

TL;DR

The paper introduces VADA and DIRT-T, two models that leverage the cluster assumption for unsupervised domain adaptation, achieving state-of-the-art results across multiple visual and non-visual benchmarks.

ABSTRACT

Domain adaptation refers to the problem of leveraging labeled data in a source domain to learn an accurate model in a target domain where labels are scarce or unavailable. A recent approach for finding a common representation of the two domains is via domain adversarial training (Ganin & Lempitsky, 2015), which attempts to induce a feature extractor that matches the source and target feature distributions in some feature space. However, domain adversarial training faces two critical limitations: 1) if the feature extraction function has high-capacity, then feature distribution matching is a weak constraint, 2) in non-conservative domain adaptation (where no single classifier can perform well in both the source and target domains), training the model to do well on the source domain hurts performance on the target domain. In this paper, we address these issues through the lens of the cluster assumption, i.e., decision boundaries should not cross high-density data regions. We propose two novel and related models: 1) the Virtual Adversarial Domain Adaptation (VADA) model, which combines domain adversarial training with a penalty term that punishes the violation the cluster assumption; 2) the Decision-boundary Iterative Refinement Training with a Teacher (DIRT-T) model, which takes the VADA model as initialization and employs natural gradient steps to further minimize the cluster assumption violation. Extensive empirical results demonstrate that the combination of these two models significantly improve the state-of-the-art performance on the digit, traffic sign, and Wi-Fi recognition domain adaptation benchmarks.

Motivation & Objective

• Address covariate shift when target labels are unavailable and a single classifier may not generalize to both source and target domains.
• Incorporate cluster-assumption-based constraints into domain adaptation to tighten the hypothesis space beyond domain adversarial training.
• Develop a refinement procedure (DIRT-T) that further minimizes target-domain cluster-assumption violations.
• Demonstrate empirical gains across diverse tasks including digits, traffic signs, general objects, and Wi-Fi activity recognition.

Proposed method

• Propose Virtual Adversarial Domain Adaptation (VADA) which combines domain adversarial training with a penalty for cluster-assumption violations and a conditional entropy loss on target data.
• Incorporate a locally-Lipschitz constraint via Virtual Adversarial Training (VAT) to stabilize entropy minimization.
• Init with VADA and perform Decision-boundary Iterative Refinement Training (DIRT-T) to further minimize target-domain cluster violations using natural gradient steps.
• DIRT-T uses a teacher-student framework where the previous model acts as a teacher to constrain updates while reducing target-side cluster violations.
• Provide an interpretation of DIRT-T as a sequence of non-conservative domain adaptation steps where domain adversarial training becomes unnecessary when refining purely on the target domain.

Experimental results

Research questions

• RQ1 Does enforcing the cluster assumption via conditional entropy minimization improve domain adaptation beyond domain adversarial training?
• RQ2 Can a refinement procedure that minimizes target-domain cluster violation further improve performance after an initial VADA model?
• RQ3 How do VAT and locally-Lipschitz constraints affect the stability and effectiveness of entropy-based clustering in domain adaptation?
• RQ4 What are the empirical gains of VADA and DIRT-T across diverse domain adaptation benchmarks (digits, traffic signs, CIFAR/STL-10, Wi-Fi)?

Key findings

• VADA improves over prior domain-adversarial methods by incorporating cluster-assumption penalties and VAT, yielding strong improvements on multiple benchmarks.
• DIRT-T consistently enhances VADA by further reducing target-domain cluster-violation, achieving state-of-the-art results across most tasks.
• DIRT-T outperforms ATT in challenging MNIST to SVHN adaptation by a substantial margin in several settings.
• Natural-gradient steps in DIRT-T contribute to stable and effective refinement of decision boundaries.
• Across Wi-Fi activity recognition, VADA significantly surpasses Source-Only and DANN; DIRT-T may not always yield gains when target clustering is already strong.