Contrastive Active Learning Under Class Distribution Mismatch.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Active learning (AL) assumes consistent data distributions, but performance degrades significantly with class distribution mismatch.
• Unlabeled data often contains instances outside the labeled data's distribution, a practical challenge in real-world applications.

Purpose of the Study:

• To address the performance degradation of active learning under class distribution mismatch.
• To formally define and decompose active learning error into valid and invalid query errors.

Main Methods:

• Proposed a contrastive active learning (ConAL) framework to learn instance semantics and distinctiveness.
• Utilized contrastive techniques to simultaneously reduce invalid query error (unknown categories) and valid query error (less informative known categories).
• Provided theoretical guarantees by proving a tight upper bound for ConAL's active learning error.

Main Results:

• ConAL demonstrated superior performance on CIFAR10 and CIFAR100 benchmark datasets.
• Achieved excellent results in cross-dataset experiments with varying class distributions.
• Validated ConAL's effectiveness on a realistic dataset, showcasing its practical applicability.

Conclusions:

• ConAL is the first active learning framework designed to handle class distribution mismatch.
• The proposed method effectively mitigates errors arising from distribution shifts, enhancing active learning robustness.