Reproduce “Interpretable epistemic uncertainty decomposition in sequential generative models via polynomial chaos surrogates” 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
Ramón Nartallo-Kaluarachchi, Shashanka Ubaru, Małgorzata J Zimoń, Dongsung Huh, Robert Manson-Sawko, Lior Horesh, Yoshua Bengio
The authors fit a polynomial chaos expansion (PCE) to generative flow network ensembles, mapping uncertain rewards to action probabilities and attributing policy variation with Sobol indices. In Buchwald–Hartwig optimization, policy variance is 71.3 for catalyst choice and 179.2 for additive choice. Additive marginal calibration coverage is 0.77; these results concern proxy rewards, not experimental outcomes.
Key insights
- PCE surrogates provide a computationally efficient method (1,000-33,000x faster than retraining) for propagating and analyzing epistemic uncertainty in generative models, making Monte Carlo estimations and counterfactual analyses practical.
- The analytical Sobol decomposition reveals an interpretable, component-specific attribution of policy uncertainty, enabling researchers to identify which aspects of generative decisions are robust or fragile to reward function variations.
- Epistemic uncertainty can lead to complex, non-scalar effects on policy, such as bifurcations in predicted distributions, highlighting critical regions where additional data would most effectively resolve structural ambiguities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.