Reproduce “Sourcerer: Sample-based Maximum Entropy Source Distribution Estimation” 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
Julius Vetter, Guy Moss, Cornelius Schröder, Richard Gao, Jakob H. Macke
Sourcerer estimates source distributions over simulator parameters by maximizing entropy while matching simulated outputs to observations with a sample-based discrepancy. It requires a differentiable simulator or surrogate, but not an evaluable likelihood. Across benchmarks, regularization raised source entropy while output fidelity remained near chance. A neuron application matched five summary statistics, but results do not establish true-source recovery or universal fidelity.
Key insights
- The maximum entropy principle offers a principled and reproducible method to resolve the ill-posedness of source distribution estimation, ensuring a unique solution that retains maximum uncertainty consistent with data.
- Leveraging sample-based distance metrics and entropy estimators bypasses the need for explicit likelihoods, significantly expanding the applicability of source distribution estimation to a wide range of complex scientific simulators.
- Maximizing entropy directly leads to the discovery of broader, yet equally valid, parameter spaces, providing a more comprehensive and less biased characterization of model uncertainty compared to methods that implicitly restrict parameter diversity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.