[Paper Review] Conditional Adversarial Domain Adaptation

Mingsheng Long, Zhangjie Cao+2 more|arXiv (Cornell University)|May 26, 2017

Domain Adaptation and Few-Shot LearningComputer Science1,428 citations

TL;DR

CDAN introduces a conditional adversarial domain discriminator that uses the cross-covariance between features and classifier predictions (with an entropy conditioning option) to align multimodal domain distributions, achieving state-of-the-art results on several benchmarks.

ABSTRACT

Adversarial learning has been embedded into deep networks to learn disentangled and transferable representations for domain adaptation. Existing adversarial domain adaptation methods may not effectively align different domains of multimodal distributions native in classification problems. In this paper, we present conditional adversarial domain adaptation, a principled framework that conditions the adversarial adaptation models on discriminative information conveyed in the classifier predictions. Conditional domain adversarial networks (CDANs) are designed with two novel conditioning strategies: multilinear conditioning that captures the cross-covariance between feature representations and classifier predictions to improve the discriminability, and entropy conditioning that controls the uncertainty of classifier predictions to guarantee the transferability. With theoretical guarantees and a few lines of codes, the approach has exceeded state-of-the-art results on five datasets.

Motivation & Objective

• Motivate robust domain adaptation when domain distributions are multimodal and diverse.
• Propose a conditional adversarial framework that leverages classifier predictions to guide domain alignment.
• Introduce multilinear and entropy conditioning to model cross-domain feature-class dependencies and uncertainty.
• Provide theoretical generalization guarantees for CDAN under domain adaptation theory.
• Demonstrate empirical improvements over state-of-the-art methods on multiple benchmark datasets.

Proposed method

• Define conditional domain adversarial network (CDAN) as a minimax game with a domain discriminator D conditioned on joint feature-class representation h=(f,g).
• Use a multilinear conditioning map f ⊗ g to capture cross-covariance between features and predictions, with a randomized approximation T⊙ for high dimensionality.
• Optionally apply entropy conditioning to weight examples by g's prediction uncertainty (CDAN+E).
• Prove that the conditioning helps model multimodal distributions and bound target risk via Δ-distance-based analysis.
• Provide a linear-time training scheme using back-propagation and a progressive training schedule for the discriminator.

Experimental results

Research questions

• RQ1 How can adversarial domain adaptation be improved to handle multimodal and complex domain shifts?
• RQ2 Does conditioning the domain discriminator on classifier predictions (and their uncertainty) yield better domain alignment and target performance?
• RQ3 Can a theoretical generalization bound be established for CDAN within domain adaptation theory?
• RQ4 What empirical gains do multilinear and entropy conditioning provide across standard domain adaptation benchmarks?

Key findings

• CDAN exceeds state-of-the-art results on five benchmark datasets.
• Multilinear conditioning captures cross-covariance between features and predictions to better model multimodal distributions.
• Entropy conditioning prioritizes easy-to-transfer examples, improving transferability (CDAN+E).
• Experiments on Office-31 and other datasets show substantial accuracy gains over baselines like DANN, JAN, and ADDA.
• The paper provides a generalization error analysis relating domain discriminator training to a Δ-distance bound between joint distributions.