[論文レビュー] Open-World Semi-Supervised Learning

Kaidi Cao, Maria Brbić+1 more|arXiv (Cornell University)|Feb 6, 2021

Domain Adaptation and Few-Shot Learning参考文献 42被引用数 48

ひとこと要約

ORCAはオープンワールド半教師あり学習のエンドツーエンド深層学習フレームワークで、ラベルなしデータの新規クラスを発見しつつ、見られたクラスを正確に分類し、見られたクラスと新規クラス間の内部クラス分散を平衡させる不確実性適応マージンを用いる。

ABSTRACT

A fundamental limitation of applying semi-supervised learning in real-world settings is the assumption that unlabeled test data contains only classes previously encountered in the labeled training data. However, this assumption rarely holds for data in-the-wild, where instances belonging to novel classes may appear at testing time. Here, we introduce a novel open-world semi-supervised learning setting that formalizes the notion that novel classes may appear in the unlabeled test data. In this novel setting, the goal is to solve the class distribution mismatch between labeled and unlabeled data, where at the test time every input instance either needs to be classified into one of the existing classes or a new unseen class needs to be initialized. To tackle this challenging problem, we propose ORCA, an end-to-end deep learning approach that introduces uncertainty adaptive margin mechanism to circumvent the bias towards seen classes caused by learning discriminative features for seen classes faster than for the novel classes. In this way, ORCA reduces the gap between intra-class variance of seen with respect to novel classes. Experiments on image classification datasets and a single-cell annotation dataset demonstrate that ORCA consistently outperforms alternative baselines, achieving 25% improvement on seen and 96% improvement on novel classes of the ImageNet dataset.

研究の動機と目的

• Generalize semi-supervised learning to open-world settings where unlabeled data contains seen and unseen classes.
• Develop an end-to-end method that can discover novel classes without knowing their number in advance.
• Mitigate bias toward seen classes by controlling intra-class variance via an uncertainty-based adaptive margin.
• Enable joint classification of seen classes and clustering into novel classes through additional classification heads.
• Demonstrate robustness to varying distributions, unknown novel-class counts, and different pretraining strategies.

提案手法

• Introduce ORCA, an end-to-end framework with an embedding backbone and multiple classification heads (one per seen class and additional heads for potential novel classes).
• Employ an uncertainty adaptive margin that modulates the supervised loss to balance learning speed between seen and novel classes, using uncertainty estimates from unlabeled data.
• Define a three-term objective: supervised loss with adaptive margin, a pairwise objective that generates pseudo-labels from unlabeled data, and a regularization term to prevent collapse to a single class.
• Use a pairwise similarity objective focusing on positive pairs to encourage same-class grouping, with pseudo-labels derived from nearest neighbors in the feature space.
• Regularize the label distribution with KL-divergence to a prior or maximum-entropy to avoid degenerate solutions.
• Optionally leverage self-supervised pretraining (SimCLR) to improve representations prior to open-world SSL training.

実験結果

リサーチクエスチョン

• RQ1 Can open-world semi-supervised learning effectively classify seen classes while discovering and clustering unseen classes from unlabeled data?
• RQ2 How can uncertainty be used to adapt margins during training to reduce intra-class variance gaps between seen and novel classes?
• RQ3 Does an end-to-end framework with multiple heads and pseudo-label guidance outperform ad hoc robust SSL or conventional novel class discovery baselines in open-world SSL?
• RQ4 Is ORCA robust to unknown numbers of novel classes and varying distributions between seen and novel data, including unbalanced datasets and different pretraining regimes?

主な発見

• ORCA consistently outperforms extended baselines across datasets, achieving notable improvements on seen and novel classes (e.g., 25% improvement on seen and 96% on novel classes for ImageNet).
• The uncertainty adaptive margin effectively reduces the gap in intra-class variance between seen and novel classes, improving pseudo-label quality and overall accuracy.
• ORCA with unknown novel-class counts can prune unused heads and still achieve strong performance, showing robustness to the number of novel classes.
• Ablation studies show that removing the adaptive margin or regularization harms performance, confirming the importance of each component.
• Self-supervised pretraining (SimCLR) can be used to boost ORCA performance, though results are reported with and without pretraining.