ENFR recovers absolute Bayesian evidence using flow-matched references
Original title: ENFR: Evidence Networks with Flow-Matched References for Amortized Simulation-Based Model Inference
Authors report better absolute-evidence estimates than a flow baseline and top density-estimation results on their benchmarks.
ENFR learns Bayes factors against a flow-matched reference and recovers absolute model evidence for Bayesian inference and goodness-of-fit tests. The authors say it improves estimation of the evidence over a flow-based baseline and performs best among density-estimation methods in their benchmarks. They also introduce, with a proof, an architecture that jointly learns evidences for a set of models and test scaling on scientific data.
Unconfirmed: Benchmark gains are author-reported and the specific scientific datasets are not named in the abstract.
- Absolute evidence can serve as a test statistic for model misspecification.
- Prior learned methods largely target Bayes factors, which are evidence ratios.
- The method builds on the Evidence Network framework of Jeffrey and Wandelt (2024).
- A joint architecture learns individual evidences for a set of models.
From the original article
1 min read at arxiv.orgThe evidence, $\mathcal{Z} = p(x | \mathcal{M})$, of a data-generating model $\mathcal{M}$ is a powerful tool in Bayesian frameworks: it enables model inference, model misspecification tests, and goodness-of-fit assessment as an alternative to standard discrepancy statistics, e. g. $χ^2$.
By Kai Lehman · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.