[Paper Review] Open Set Domain Adaptation By Novel Class Discovery

Jingyu Zhuang, Ziliang Chen+3 more|arXiv (Cornell University)|Mar 7, 2022

Domain Adaptation and Few-Shot Learning8 citations

TL;DR

This paper proposes Self-supervised Class-Discovering Adapter (SCDA), a novel approach for Open Set Domain Adaptation (OSDA) that discovers implicit novel classes in target data instead of treating them as a single 'unknown' class. SCDA alternately performs implicit class discovery and domain-adaptive feature learning, achieving state-of-the-art performance on OSDA benchmarks and highly accurate novel class estimation.

ABSTRACT

In Open Set Domain Adaptation (OSDA), large amounts of target samples are drawn from the implicit categories that never appear in the source domain. Due to the lack of their specific belonging, existing methods indiscriminately regard them as a single class unknown. We challenge this broadly-adopted practice that may arouse unexpected detrimental effects because the decision boundaries between the implicit categories have been fully ignored. Instead, we propose Self-supervised Class-Discovering Adapter (SCDA) that attempts to achieve OSDA by gradually discovering those implicit classes, then incorporating them to restructure the classifier and update the domain-adaptive features iteratively. SCDA performs two alternate steps to achieve implicit class discovery and self-supervised OSDA, respectively. By jointly optimizing for two tasks, SCDA achieves the state-of-the-art in OSDA and shows a competitive performance to unearth the implicit target classes.

Motivation & Objective

• Challenge the common practice of treating all out-of-distribution target samples as a single 'unknown' class in OSDA.
• Address the detrimental effect of ignoring inter-class structure among implicit target categories, which causes feature confusion under the cluster assumption.
• Develop a self-supervised framework that discovers the number and distribution of implicit classes in unlabeled target data.
• Improve OSDA performance by iteratively restructuring the classifier and updating features based on discovered implicit classes.
• Demonstrate that discovering the internal structure of unknown target samples leads to better generalization and reduced performance degradation near decision boundaries.

Proposed method

• Apply adversarial pre-training to initially separate unknown target samples from known classes.
• Design an unsupervised algorithm for implicit class discovery that estimates the number of novel classes and clusters unknown samples accordingly.
• Restructure the classifier to treat discovered implicit classes as distinct categories rather than a single 'unknown' class.
• Alternate between implicit class discovery and model fine-tuning to iteratively improve feature disentanglement and classification accuracy.
• Leverage self-supervised pseudo-labeling from discovered classes to update the model without requiring ground-truth labels.
• Use iterative refinement to enhance both class discovery accuracy and domain adaptation performance.

Experimental results

Research questions

• RQ1 Can discovering the internal structure of unknown target samples improve OSDA performance compared to treating them as a single 'unknown' class?
• RQ2 How accurately can a self-supervised method estimate the number of implicit novel classes in target domains?
• RQ3 Does iterative refinement of discovered classes lead to better feature disentanglement and classification performance?
• RQ4 How does SCDA perform under real-world noisy and imbalanced target domain settings?
• RQ5 What is the relative contribution of accurate class number estimation versus iterative model optimization in achieving state-of-the-art OSDA?

Key findings

• SCDA achieves state-of-the-art performance on three OSDA benchmarks, with a 55.3% average OS accuracy on OfficeHome to DomainNet transfers, outperforming prior SoTA methods like SHOT (54.3%) and STA (48.2%).
• On the Office-31 dataset, SCDA estimates the number of unknown classes with an average error of only 0.9, significantly outperforming DTC (6.1) and SC (4.2).
• In the challenging DomainNet⋆ benchmarks with 297 ground-truth unknown classes, SCDA estimates the number with a mean error of 58.5, demonstrating robustness in high-complexity settings.
• Ablation studies show that discovering implicit classes improves performance over treating them as 'unknown', and iterative refinement with estimated class numbers achieves performance close to that with ground-truth class counts.
• SCDA shows faster convergence and higher upper-bound performance in real-world noisy and imbalanced scenarios, such as OfficeHome to DomainNet⋆ transfers.
• Visualization and analysis confirm that SCDA reduces feature confusion between known and unknown classes, especially near decision boundaries, by preserving inter-class structure in the unknown.