Explorative Latent Self-Supervised Active Search Algorithm (ELSA)
Primary Area: representation learning for computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Computer Vision, Active Learning, Interactive Labeling, Self-Supervised Learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: In computer vision, attaining exceptional performance often necessitates access to large labeled datasets. The creation of extensive datasets through manual annotation is not only cost-prohibitive but also practically infeasible due to the scarcity of positive samples in imbalanced datasets where negative samples dominate. To tackle this intricate problem, we introduce Efficient Latent Space-based Self-Supervised Active Learning Search (ELSA), an active learning-based labeling assistant. ELSA distinguishes itself from existing interactive annotation methods by focusing exclusively on positive class labeling in massively imbalanced datasets replete with a substantial number of negative samples. Through the automatic exclusion of the majority of negative samples, ELSA achieves a remarkable level of precision and accuracy in its search. This novel framework comprises three fundamental components: a)an iterative Nearest Neighbor Search, b)a Sophisticated Random Sampler, c)a Linear Head powered by Active Learning. Our comprehensive study provides insights into the interplay of these components and their collective impact on search efficiency. Notably, we demonstrate that ELSA achieves orders of magnitude superior performance, in average starting with as little as 5 or less positive samples in ImageNet 1k we managed to detect as much as 80\% of all the examples belonging to that class by only labeling as little as 0.67\% of the entire dataset manually.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
Supplementary Material: zip
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 9259
Loading