Reproduce “Certainty Pooling for Multiple Instance Learning” 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
Jacob Gildenblat, Ido Ben-Shaul, Zvi Lapp, Eldad Klaiman
Certainty Pooling, a novel Multiple Instance Learning (MIL) pooling function, leverages Monte-Carlo dropout to estimate prediction certainty for individual instances, improving bag-level and instance-level classification. This approach demonstrates superior performance on low evidence ratio datasets, achieving higher AUCs and enhanced interpretability on both a synthetic MNIST-Bags dataset and the Camelyon16 lymph node metastasis detection challenge.
Key insights
- Weighting instance predictions by their model certainty provides a robust mechanism to reduce the impact of unstable or unreliable instance outputs, leading to more stable bag-level decisions.
- Incorporating uncertainty directly into the pooling mechanism, rather than learning attention weights, yields a more interpretable MIL model where higher-confidence instances drive the prediction.
- By providing more meaningful gradients, Certainty Pooling helps the model learn a more effective instance-level classifier, improving its ability to distinguish positive from negative instances within bags.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.