Active Learning via Classifier Impact and Greedy Selection for Interactive Image Retrieval

TMLR Paper1428 Authors

Abstract:

Active Learning (AL) is a user-interactive approach aimed at reducing annotation costs by selecting the most crucial examples to label. Although AL has been extensively studied for image classification tasks, the specific scenario of interactive image retrieval has received
relatively little attention. This scenario presents unique characteristics, including an open set and class-imbalanced binary classification, starting with very few labeled samples. To address this specific scenario, we introduce a novel batch-mode Active Learning framework named GAL (Greedy Active Learning) that incorporates a new acquisition function for sample selection that measures the impact of each unlabeled sample on the classifier. We further embed this strategy in a greedy selection approach. We evaluate our framework with both linear (SVM) and non-linear (Gaussian Process, MLP) classifiers. For SVM and MLP, our method considers a pseudo-label strategy for each sample while ensuring tractability through a greedy approach. Considering our Gaussian Process acquisition function, we show a theoretical guarantee for the greedy approximation. Finally, we assess our performance on the interactive content-based image retrieval task and demonstrate its superiority over existing approaches and common baselines.

Submission Length: Long submission (more than 12 pages of main content)

Changes Since Last Submission: Please find attached the revised manuscript, wherein we have carefully incorporated all the review comments. Notably, we have color-coded the changes in this version to facilitate easy tracking. The updated manuscript incorporates new experiments and comparisons as requested by the reviewers, along with clarifications and modifications in response to their valuable feedback. We express sincere gratitude for the constructive comments from all reviewers, which have significantly contributed to enhancing the quality of the manuscript.

Assigned Action Editor: ~ERIC_EATON1

Submission Number: 1428