Reproduce “Multiple Instance Learning: A Survey of Problem Characteristics and Applications” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Marc-André Carbonneau, Veronika Cheplygina, Eric Granger, Ghyslain Gagnon
The paper presents the first comprehensive survey of Multiple Instance Learning (MIL) by formally identifying and categorizing its problem characteristics into four distinct groups. It empirically confirms that MIL algorithm performance is highly dependent on these characteristics, demonstrating inconsistencies between instance-level and bag-level optimization and highlighting the impact of witness rate and negative distribution representativeness.
Key insights
- The optimal algorithm for bag classification often differs from that for instance classification, highlighting that these two objectives require distinct algorithmic approaches.
- Lower witness rates consistently lead to decreased accuracy across all MIL methods for both instance and bag classification, emphasizing the fundamental challenge of sparse positive instances.
- Most algorithms show performance degradation when test negative instance distributions diverge from training distributions, though some methods like C-kNN can surprisingly improve under such conditions due to their distance-based mechanisms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.