Fundamental Linear Algebra Problem of Gaussian Inference — mentioned by shoo

Tue, 31 Aug 2021 14:27:09 GMT

> Theoretical grounds for applying the same update rules to loopy graphs were later developed that explain loopy BP as an approximate variational inference method in which inference is cast as an optimization problem. Instead of directly minimizing the factor energies (as is done in MAP inference), loopy BP minimizes the KL divergence between the posterior and a variational distribution which we use as a proxy for the marginals after optimization > As the Bethe free energy is non-convex, loopy BP is not guaranteed to converge and even when it does it may converge to the wrong marginals.…

Paper and discussion history. A mention is not an endorsement.

Fundamental Linear Algebra Problem of Gaussian Inference · HackerPapers discussions

Fundamental Linear Algebra Problem of Gaussian Inference — mentioned by shoo