[论文解读] Deep Cocktail Network: Multi-source Unsupervised Domain Adaptation with Category Shift

Ruijia Xu, Ziliang Chen+3 more|arXiv (Cornell University)|Mar 2, 2018

Domain Adaptation and Few-Shot Learning参考文献 46被引用 54

一句话总结

Introduces Deep Cocktail Network (DCTN) for multi-source unsupervised domain adaptation with possible category shift, using multi-way adversarial learning and source-wise route weighting to fuse predictions for the target domain.

ABSTRACT

Unsupervised domain adaptation (UDA) conventionally assumes labeled source samples coming from a single underlying source distribution. Whereas in practical scenario, labeled data are typically collected from diverse sources. The multiple sources are different not only from the target but also from each other, thus, domain adaptater should not be modeled in the same way. Moreover, those sources may not completely share their categories, which further brings a new transfer challenge called category shift. In this paper, we propose a deep cocktail network (DCTN) to battle the domain and category shifts among multiple sources. Motivated by the theoretical results in \cite{mansour2009domain}, the target distribution can be represented as the weighted combination of source distributions, and, the multi-source unsupervised domain adaptation via DCTN is then performed as two alternating steps: i) It deploys multi-way adversarial learning to minimize the discrepancy between the target and each of the multiple source domains, which also obtains the source-specific perplexity scores to denote the possibilities that a target sample belongs to different source domains. ii) The multi-source category classifiers are integrated with the perplexity scores to classify target sample, and the pseudo-labeled target samples together with source samples are utilized to update the multi-source category classifier and the feature extractor. We evaluate DCTN in three domain adaptation benchmarks, which clearly demonstrate the superiority of our framework.

研究动机与目标

• Motivate and address unsupervised domain adaptation when multiple diverse sources are available and may not share the same categories (category shift).
• Propose a deep cocktail network (DCTN) that represents the target as a weighted combination of source distributions, and optimize via alternating adversarial and discriminative steps.
• Develop a practical training pipeline that includes multi-way adversarial domain adaptation and target discriminative adaptation with pseudo-labels for the target data.

提出的方法

• Propose four components: a shared feature extractor, a multi-source domain discriminator, a multi-source category classifier, and a non-learnable target classification operator.
• Use multi-way adversarial learning to minimize target-source discrepancies and generate source-specific perplexity scores.
• Compute target predictions by weighting source-specific classifiers according to perplexity scores and aggregate to assign target labels (category shift compatible).
• Train with an alternating protocol: (i) adversarial/domain alignment steps; (ii) discriminative adaptation with pseudo-labeled targets and source data.
• Present a learning algorithm that includes online hard domain batch mining to focus on the most informative source-target pairs.

实验结果

研究问题

• RQ1 How can we effectively perform unsupervised domain adaptation when multiple sources with potentially different category sets (category shift) are available?
• RQ2 Can a distribution-weighted combination of source classifiers improve target prediction in multi-source domain adaptation?
• RQ3 Does incorporating target pseudo-labels with multi-source classifiers enhance discriminative transfer under category shift?
• RQ4 What is the impact of multi-way adversarial learning and source-specific perplexity weighting on target performance across benchmarks?

主要发现

Standards	Models	A,W→D	A,D→W	D,W→A	Avg
Office-31 vanilla	DCTN (ours)	99.6	96.9	54.9	83.8
Office-31 vanilla	RTN	99.6?	96.8?	51.0?	73.7?
Office-31 vanilla	DAN	99.0	96.0	54.0	72.9
Office-31 vanilla	RevGrad	99.2	96.4	53.4	74.3
Office-31 vanilla	DRCN	99.0	96.4	56.0	73.6
Office-31 vanilla	GFK	95.0	95.6	52.4	68.7
Office-31 vanilla	TCA	95.2	93.2	51.6	68.8
Office-31 vanilla	Source combine ( Source only )	98.1	93.2	50.2	80.5
Office-31 vanilla	RevGrad (second)	98.8	96.2	54.6	83.2
Office-31 vanilla	DAN (second)	98.8	95.2	53.4	82.5
Office-31 vanilla	Multi-source (Source only)	98.2	92.7	51.6	80.8
Digits-five vanilla	DCTN (ours)	77.5	70.9	74.2	74.2

• DCTN achieves state-of-the-art performance on Office-31, ImageCLEF-DA, and Digits-five benchmarks in vanilla MDA settings.
• In category shift settings, DCTN maintains robust performance and shows positive transfer gains, while several baselines suffer degradation.
• Across Office-31 (vanilla), DCTN attains 99.6, 96.9, 54.9, 83.8 (A,W→D; A,D→W; D,W→A; Avg).
• Across ImageCLEF-DA (vanilla), DCTN achieves 68.8, 90.0, 83.5 (I,C→P; I,P→C; P,C→I) with an 80.8 avg.
• Across Digits-five (vanilla), DCTN shows strong improvement over baselines in the reported domain transfers (e.g., 77.5, 70.9, 74.2 on mm, mt, sy, up→sv and mt, sv, sy, up→mm).
• Ablation analysis indicates both the multi-way adversary and the pseudo-labeling strategy contribute to performance gains.