Reproduce “Multidimensional Uncertainty Quantification via Optimal Transport” 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
Nikita Kotelevskii, Maiya Goloburda, Vladimir Kondratyev, Alexander Fishkov, Mohsen Guizani, Eric Moulines, Maxim Panov
The paper presents a method that combines multiple uncertainty scores into a single ranking using entropy-regularized optimal transport. It maps score vectors from unlabeled calibration inputs to a reference distribution and ranks new inputs by distance from its center. In a synthetic logistic-regression test, it achieved perfect out-of-distribution detection while retaining 0.710 misclassification-detection ROC-AUC, but did not lead both tasks.
Key insights
- A multidimensional representation of uncertainty, created by combining diverse measures, offers a more comprehensive and robust view of model reliability than single-scalar approaches.
- Optimal Transport provides a theoretically grounded, non-parametric mechanism to establish a principled ordering for complex, multivariate uncertainty signals.
- The framework acts as a 'label-free calibration layer,' intelligently fusing multiple signals to simplify UQ integration into machine learning workflows for various downstream tasks without needing task-specific re-tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.