[Paper Review] Implicit Class-Conditioned Domain Alignment for Unsupervised Domain Adaptation

Xiang Jiang, Qicheng Lao+2 more|arXiv (Cornell University)|Jun 9, 2020

Domain Adaptation and Few-Shot Learning52 citations

TL;DR

Proposes an implicit, sampling-based class-conditioned domain alignment to improve unsupervised domain adaptation, addressing within-domain class imbalance and between-domain distribution shift without relying on pseudo-label optimization.

ABSTRACT

We present an approach for unsupervised domain adaptation---with a strong focus on practical considerations of within-domain class imbalance and between-domain class distribution shift---from a class-conditioned domain alignment perspective. Current methods for class-conditioned domain alignment aim to explicitly minimize a loss function based on pseudo-label estimations of the target domain. However, these methods suffer from pseudo-label bias in the form of error accumulation. We propose a method that removes the need for explicit optimization of model parameters from pseudo-labels directly. Instead, we present a sampling-based implicit alignment approach, where the sample selection procedure is implicitly guided by the pseudo-labels. Theoretical analysis reveals the existence of a domain-discriminator shortcut in misaligned classes, which is addressed by the proposed implicit alignment approach to facilitate domain-adversarial learning. Empirical results and ablation studies confirm the effectiveness of the proposed approach, especially in the presence of within-domain class imbalance and between-domain class distribution shift.

Motivation & Objective

• Address the challenge of class imbalance within domains and distribution shift between domains in unsupervised domain adaptation (UDA).
• Eliminate reliance on explicit pseudo-label optimization for class-conditioned alignment.
• Provide theoretical insight into domain-discriminator shortcuts arising from class misalignment and propose a sampling-based remedy.
• Demonstrate robustness and improvements across standard UDA benchmarks under extreme class distribution scenarios.

Proposed method

• Decompose empirical HΔH divergence into class-aligned and class-misaligned components to show the harmful effect of misalignment.
• Introduce implicit class-conditioned domain alignment via class-aligned minibatch sampling driven by pseudo-labels, with a fixed target label distribution p(y).
• Sample S: x ~ pS(x|y)p(y) and T: x ~ pT(x|ŷ)p(y) to create class-aligned minibatches for training.
• Integrate implicit sampling with adversarial domain alignment (e.g., MDD) and apply a masking scheme to the domain discrepancy estimator to mitigate the class-misalignment shortcut.
• Provide an oracle and a pseudo-label-based version to study robustness to pseudo-label errors.
• Show that the method is orthogonal to the choice of domain adaptation algorithm and yields improvements on multiple benchmarks.

Experimental results

Research questions

• RQ1 Can sampling-based implicit class-conditioned alignment reduce misalignment between class distributions across domains without explicit pseudo-label optimization?
• RQ2 How does addressing class misalignment affect empirical domain divergence estimates and domain-invariant representation learning?
• RQ3 Is implicit alignment robust to pseudo-label errors and extreme within-domain/class-distribution shift across standard UDA benchmarks?
• RQ4 Does the approach synergize with existing domain discrepancy measures (e.g., MDD) and improve performance over explicit class-conditioned alignment?

Key findings

• Implicit alignment achieves state-of-the-art results under extreme within-domain class imbalance and between-domain distribution shift (Office-Home RS-UT).
• Compared to explicit alignment, implicit alignment is more robust to pseudo-label inaccuracies and reduces error accumulation.
• Across Office-Home, Office-31, and VisDA2017, implicit alignment consistently improves upon strong baselines, including MDD and several explicit methods.
• A masking scheme on the domain discrepancy estimator, combined with class-aligned sampling, effectively mitigates the domain-discriminator shortcut in misaligned classes.
• Ablation studies show both input-space sampling and divergence-metric masking are essential for the observed gains.
• When integrated with DANN in synthetic experiments, implicit alignment further improves accuracy under various imbalance scenarios.