[Paper Review] Multi-Adversarial Domain Adaptation

Zhongyi Pei, Zhangjie Cao+2 more|arXiv (Cornell University)|Sep 4, 2018

Domain Adaptation and Few-Shot LearningComputer Science35 references89 citations

TL;DR

MADA introduces multiple class-wise domain discriminators to perform fine-grained, multi-adversarial alignment of source and target data, improving unsupervised domain adaptation over prior single-discriminator methods.

ABSTRACT

Recent advances in deep domain adaptation reveal that adversarial learning can be embedded into deep networks to learn transferable features that reduce distribution discrepancy between the source and target domains. Existing domain adversarial adaptation methods based on single domain discriminator only align the source and target data distributions without exploiting the complex multimode structures. In this paper, we present a multi-adversarial domain adaptation (MADA) approach, which captures multimode structures to enable fine-grained alignment of different data distributions based on multiple domain discriminators. The adaptation can be achieved by stochastic gradient descent with the gradients computed by back-propagation in linear-time. Empirical evidence demonstrates that the proposed model outperforms state of the art methods on standard domain adaptation datasets.

Motivation & Objective

• Motivate and address negative transfer in domain adaptation by exploiting multimode structures in data distributions.
• Propose a framework that uses multiple class-wise domain discriminators to enable fine-grained alignment between source and target domains.
• Demonstrate that weighting domain discriminator inputs by label-prediction probabilities improves transfer performance.

Proposed method

• Split the single domain discriminator into K class-wise discriminators G_d^k, each responsible for aligning data belonging to class k.
• Use the label predictor G_y to generate a per-sample class probability ŷ_i^k to weight the input to each G_d^k, creating a soft, attention-like mechanism.
• Optimize a joint objective that minimizes source classification loss plus a weighted sum of domain-discriminator losses across all K discriminators.
• Train with stochastic gradient descent and a gradient-reversal mechanism to promote domain confusion while preserving discriminative class structure.
• Provide linear-time back-propagation for efficient training compatible with deep networks.

Experimental results

Research questions

• RQ1 Can leveraging multimode (class-wise) structures through multiple discriminators improve alignment between source and target domains?
• RQ2 Does weighting domain-discriminator inputs by class-probabilities reduce negative transfer compared to a single discriminator?
• RQ3 How does MADA perform relative to existing deep-domain adaptation methods on standard benchmarks?

Key findings

• MADA outperforms state-of-the-art methods on Office-31 and ImageCLEF-DA benchmarks in unsupervised domain adaptation.
• MADA shows substantial gains on hard transfer tasks where domain discrepancy is large (e.g., A→W, A→D, D→A, W→A).
• In ablations and robustness analyses, MADA reduces negative transfer better than single-discriminator approaches and maintains stable convergence.
• Feature visualizations indicate MADA yields more discriminative and domain-invariant features than RevGrad.
• Sharing strategies that fully share parameters across multiple discriminators can degrade performance, supporting the need for distinct class-wise discriminators.
• Proxy A-distance analyses show MADA features yield smaller domain discrepancy than RevGrad or plain ResNet features.