Author name: Pavel Prochazka

arXiv:2605.08446v2 Announce Type: replace
Abstract: We propose training Bayesian neural networks by directly minimizing the Bethe free energy rather than maximizing a variational lower bound. On tree-structured factor graphs the Bethe free energy is e…

arXiv:2605.08446v1 Announce Type: new
Abstract: We propose training Bayesian neural networks by directly minimizing the Bethe free energy rather than maximizing a variational lower bound. On tree-structured factor graphs the Bethe free energy is exact…