Bayesian models quantify uncertainty and facilitate optimal decision-making
in downstream applications. For most models, however, practitioners are forced
to use approximate inference techniques that lead to sub-optimal decisions due
to incorrect posterior predictive distributions. We present a novel approach
that corrects for inaccuracies in posterior inference by altering the
decision-making process. We train a separate model to make optimal decisions
under the approximate posterior, combining interpretable Bayesian modeling with
optimization of direct predictive accuracy in a principled fashion. The
solution is generally applicable as a plug-in module for predictive
decision-making for arbitrary probabilistic programs, irrespective of the
posterior inference strategy. We demonstrate the approach empirically in
several problems, confirming its potential.
Description
[1909.04919] Correcting Predictions for Approximate Bayesian Inference
%0 Journal Article
%1 kusmierczyk2019correcting
%A Kuśmierczyk, Tomasz
%A Sakaya, Joseph
%A Klami, Arto
%D 2019
%K approximate bayesian sampling uncertainty
%T Correcting Predictions for Approximate Bayesian Inference
%U http://arxiv.org/abs/1909.04919
%X Bayesian models quantify uncertainty and facilitate optimal decision-making
in downstream applications. For most models, however, practitioners are forced
to use approximate inference techniques that lead to sub-optimal decisions due
to incorrect posterior predictive distributions. We present a novel approach
that corrects for inaccuracies in posterior inference by altering the
decision-making process. We train a separate model to make optimal decisions
under the approximate posterior, combining interpretable Bayesian modeling with
optimization of direct predictive accuracy in a principled fashion. The
solution is generally applicable as a plug-in module for predictive
decision-making for arbitrary probabilistic programs, irrespective of the
posterior inference strategy. We demonstrate the approach empirically in
several problems, confirming its potential.
@article{kusmierczyk2019correcting,
abstract = {Bayesian models quantify uncertainty and facilitate optimal decision-making
in downstream applications. For most models, however, practitioners are forced
to use approximate inference techniques that lead to sub-optimal decisions due
to incorrect posterior predictive distributions. We present a novel approach
that corrects for inaccuracies in posterior inference by altering the
decision-making process. We train a separate model to make optimal decisions
under the approximate posterior, combining interpretable Bayesian modeling with
optimization of direct predictive accuracy in a principled fashion. The
solution is generally applicable as a plug-in module for predictive
decision-making for arbitrary probabilistic programs, irrespective of the
posterior inference strategy. We demonstrate the approach empirically in
several problems, confirming its potential.},
added-at = {2020-01-13T12:35:55.000+0100},
author = {Kuśmierczyk, Tomasz and Sakaya, Joseph and Klami, Arto},
biburl = {https://www.bibsonomy.org/bibtex/2e12db6bdd904d5a1b61903f2f414da96/kirk86},
description = {[1909.04919] Correcting Predictions for Approximate Bayesian Inference},
interhash = {62d524eeb35b6f473f93f8dbf5aec29d},
intrahash = {e12db6bdd904d5a1b61903f2f414da96},
keywords = {approximate bayesian sampling uncertainty},
note = {cite arxiv:1909.04919},
timestamp = {2020-01-13T12:35:55.000+0100},
title = {Correcting Predictions for Approximate Bayesian Inference},
url = {http://arxiv.org/abs/1909.04919},
year = 2019
}