[Paper Review] Multi-source Domain Adaptation in the Deep Learning Era: A Systematic Survey

Sicheng Zhao, Bo Li+3 more|arXiv (Cornell University)|Feb 26, 2020

Domain Adaptation and Few-Shot Learning61 references73 citations

TL;DR

This paper provides a comprehensive survey of multi-source domain adaptation (MDA) in deep learning, defining MDA, reviewing datasets, comparing latent-space transformation and intermediate-domain generation methods, and outlining future directions.

ABSTRACT

In many practical applications, it is often difficult and expensive to obtain enough large-scale labeled data to train deep neural networks to their full capability. Therefore, transferring the learned knowledge from a separate, labeled source domain to an unlabeled or sparsely labeled target domain becomes an appealing alternative. However, direct transfer often results in significant performance decay due to domain shift. Domain adaptation (DA) addresses this problem by minimizing the impact of domain shift between the source and target domains. Multi-source domain adaptation (MDA) is a powerful extension in which the labeled data may be collected from multiple sources with different distributions. Due to the success of DA methods and the prevalence of multi-source data, MDA has attracted increasing attention in both academia and industry. In this survey, we define various MDA strategies and summarize available datasets for evaluation. We also compare modern MDA methods in the deep learning era, including latent space transformation and intermediate domain generation. Finally, we discuss future research directions for MDA.

Motivation & Objective

• Define and categorize multi-source domain adaptation (MDA) settings and strategies.
• Summarize publicly available datasets used for evaluating MDA.
• Compare deep MDA approaches, focusing on latent-space transformation and intermediate-domain generation.
• Discuss theoretical insights and propose promising future research directions.

Proposed method

• Categorize MDA by alignment strategies: latent space transformation and intermediate domain generation.
• Review discrepancy-based and adversarial methods for latent-space alignment.
• Discuss shared versus unshared feature extractors and classifier alignment across sources.
• Describe intermediate-domain generation approaches (e.g., GAN-based) to bridge source and target domains.
• Highlight weightings and source selection strategies for combining multiple sources.

Experimental results

Research questions

• RQ1 What(problem) definitions and strategy taxonomies best capture multi-source domain adaptation in deep learning?
• RQ2 How do existing MDA methods perform across standard datasets when comparing latent-space transformation versus intermediate-domain generation?
• RQ3 What are the key theoretical and practical insights guiding when to use which MDA strategy?
• RQ4 What future directions are most promising for improving MDA in real-world, multi-source settings?

Key findings

• Directly combining multiple sources into one and applying single-source DA often underperforms the best single-source approach.
• MDA methods are commonly organized into latent-space transformation and intermediate-domain generation frameworks.
• Discrepancy-based and adversarial approaches are prevalent for latent-space alignment, with Wasserstein-based strategies offering stability advantages.
• Weighting and selecting the most relevant sources and samples improves target adaptation performance.
• Some methods employ either shared or unshared feature extractors, trading off parameter efficiency for alignment granularity.
• Intermediate-domain generation with pixel-level alignment can help improve fine-grained tasks like semantic segmentation.