Bag-level Self-supervised instance based distance in Multiple Instance Learning

Avital Rose, Yoram Louzoun

Keywords: Multiple Instance Learning, Self supervised, Bag, Instance, Energy distance, Embedding

TL;DR: We propose an embedding of instances in bags to produce a distance that maintains

Abstract:

Multiple Instance Learning (MIL) methods are typically supervised. However, a bag-to-bag metric is needed in many applications, including clustering, statistical tests, and dimension reduction.
Such a metric should differentiate between bags, regardless of the sparsity or overlap between the instances of the bags. We propose SUMIT (Self sUpervised MIL dIsTance) as an instance-embedding-based distance that maximizes the distinction between bags. SUMIT is optimized using five criteria: self-similarity within a bag, quality of instance reconstruction, robustness to sampling depth, conservation of triangle inequality, and separation of instances to clusters. We show using current standard MIL datasets and a novel wiki-based set of wiki topics that the within bag-similarity loss is the most important for a bag-to-bag metric that best separates bags of similar classes. SUMIT bridges the gap between instance-level and bag-level approaches, by keeping the embedding of all instances but ensuring their proximity within a bag.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3495