Contrastive Open-Set Active Learning-Based Sample Selection for Image Classification

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Active Learning (AL) typically assumes all unlabeled data is in-distribution (ID).
• Open-set AL scenarios involve both ID and out-of-distribution (OOD) samples in unlabeled data.
• Standard AL methods fail by selecting uncertain OOD samples, wasting computational resources and reducing model performance.

Purpose of the Study:

• To develop an effective Active Learning strategy for open-set scenarios with mixed ID and OOD unlabeled data.
• To improve the selection of informative ID samples while mitigating the misclassification of OOD samples.
• To enhance the representation learning capabilities of classifiers within the AL framework.

Main Methods:

• Introduced two novel criteria: contrastive confidence (ID possibility) and historical divergence (sample hardness).
• Developed a contrastive clustering framework that integrates OOD detection into the classifier.
• Balanced contrastive confidence and historical divergence to prioritize informative ID samples.

Main Results:

• The proposed method successfully identifies and avoids selecting OOD samples.
• Achieved state-of-the-art performance on several benchmark datasets for open-set Active Learning.
• Enhanced the network's representation learning without requiring separate OOD detection modules.

Conclusions:

• The novel approach effectively addresses the challenges of open-set Active Learning.
• The contrastive clustering framework offers a unified solution for sample selection and OOD detection.
• The proposed method demonstrates superior performance and efficiency in complex AL scenarios.