Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Deep Co-Training with Task Decomposition for Semi-Supervised Domain Adaptation (2007.12684v5)

Published 24 Jul 2020 in cs.CV and cs.LG

Abstract: Semi-supervised domain adaptation (SSDA) aims to adapt models trained from a labeled source domain to a different but related target domain, from which unlabeled data and a small set of labeled data are provided. Current methods that treat source and target supervision without distinction overlook their inherent discrepancy, resulting in a source-dominated model that has not effectively used the target supervision. In this paper, we argue that the labeled target data needs to be distinguished for effective SSDA, and propose to explicitly decompose the SSDA task into two sub-tasks: a semi-supervised learning (SSL) task in the target domain and an unsupervised domain adaptation (UDA) task across domains. By doing so, the two sub-tasks can better leverage the corresponding supervision and thus yield very different classifiers. To integrate the strengths of the two classifiers, we apply the well-established co-training framework, in which the two classifiers exchange their high confident predictions to iteratively "teach each other" so that both classifiers can excel in the target domain. We call our approach Deep Co-training with Task decomposition (DeCoTa). DeCoTa requires no adversarial training and is easy to implement. Moreover, DeCoTa is well-founded on the theoretical condition of when co-training would succeed. As a result, DeCoTa achieves state-of-the-art results on several SSDA datasets, outperforming the prior art by a notable 4% margin on DomainNet. Code is available at https://github.com/LoyoYang/DeCoTa

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (7)
  1. Luyu Yang (8 papers)
  2. Yan Wang (733 papers)
  3. Mingfei Gao (26 papers)
  4. Abhinav Shrivastava (120 papers)
  5. Kilian Q. Weinberger (105 papers)
  6. Wei-Lun Chao (92 papers)
  7. Ser-Nam Lim (116 papers)
Citations (68)

Summary

Deep Co-Training with Task Decomposition for Semi-Supervised Domain Adaptation

The paper introduces Deep Co-Training with Task Decomposition (DeCoTa), a novel approach for Semi-Supervised Domain Adaptation (SSDA) that integrates task decomposition with co-training to enhance model performance across domains. The primary objective of SSDA is to adapt models trained from a labeled source domain to a target domain, where limited labeled and abundant unlabeled data are available. The traditional methods often dominated by source data fail to effectively leverage the labeled target data due to inherent discrepancies between the source and target domains.

Approach and Methods

DeCoTa proposes an explicit task decomposition strategy to address the challenges in SSDA. The framework decomposes the SSDA task into two sub-tasks: a Semi-Supervised Learning (SSL) task within the target domain and an Unsupervised Domain Adaptation (UDA) task across domains. This decomposition allows each task to utilize its corresponding supervision optimally, producing distinct classifiers with complementary strengths.

The framework employs co-training to integrate the strengths of these two classifiers. Each classifier iteratively refines its performance by using high-confident predictions from the other classifier. This co-training mechanism eliminates the need for adversarial training, simplifying the implementation of DeCoTa.

Implementation and Theoretical Foundation

DeCoTa is grounded on theoretical conditions ensuring co-training success, specifically focusing on the relaxed ε-expandability condition that justifies task decomposition and the co-training methodology. The algorithm uses pseudo-labeling strategies informed by MixUp to improve the quality and coverage of pseudo-labels, resulting in better denoising and domain bridging.

In practice, DeCoTa leverages Mini-Batch Stochastic Gradient Descent (SGD) for concurrent training of the two classifiers. It utilizes the high-confidence pseudo-label predictions to refine the classifiers' learning iteratively. By optimizing the prediction loop for both labeled and pseudo-labeled data through MixUp, DeCoTa successfully minimizes noise and enhances model generalization across domains.

Experimental Insights and Results

Empirical evaluations were conducted on benchmark datasets such as DomainNet and Office-Home under various SSDA settings (such as one-shot and three-shot). DeCoTa achieved state-of-the-art performance, outpacing previous methods by significant margins. For instance, it demonstrated a notable 4% improvement over current approaches on the DomainNet dataset. Notably, the experiments confirmed that task decomposition aligns with co-training's theoretical conditions, enhancing prediction accuracy.

Comparisons and Ablation Studies

Comparative analyses with baseline models, including DANN, ENT, MME, and recent SSDA approaches such as UODA and APE, showcase DeCoTa's superior performance. Further ablation studies reveal the necessity of both task decomposition and co-training. Variants of the algorithm, such as MIxUp Self-Training (MiST) and two-view MiST, were evaluated to illustrate the distinct advantages of integrating task-specific views along with collaborative training.

Future Implications and Directions

DeCoTa opens new avenues for SSDA by emphasizing task decomposition principles collaboratively refined using co-training techniques. It is a significant step towards more effective domain adaptation strategies, promising greater adaptability and robustness across varying domain conditions. Future research could explore the scalability of DeCoTa to even more complex and larger domain scenarios and evaluate its performance in real-world applications where domain shifts are frequent.

In summary, the paper provides a comprehensive, theoretically justified, and empirically validated approach to SSDA that enhances model performance through strategic task decomposition and co-training. It sets a new benchmark in the field and presents a foundation for further exploration and refinement in domain adaptation challenges.

Youtube Logo Streamline Icon: https://streamlinehq.com