arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2010.08022cs.LGcs.ROmath.STstat.TH
收藏

高斯推论的基本线性代数问题

Fundamental Linear Algebra Problem of Gaussian Inference

Institute for Aerospace Studies(航空航天研究所);University of Toronto(多伦多大学)

Timothy D Barfoot

摘要:

许多贝叶斯推理技术试图将后验近似为高斯分布,这是一个基本的线性代数问题,必须解决协方差的均值和关键项。即使当真实后验不是高斯时(例如,在非线性测量函数的情况下),我们可以使用在每次迭代中重复地求解该线性代数问题的变分方案。在大多数情况下,问题不在于这个问题是否存在解决方案,而在于我们如何利用特定于问题的结构来有效地找到它。我们的贡献是清楚地说明基本线性代数问题的高斯推理(FLAPOGI),并提供一个新的介绍(使用克罗内克代数)的不那么知名的结果高桥等。(1973),使之有可能解决的关键项目的协方差矩阵。我们首先提供了一个全球性的解决方案,然后一个本地版本,可以实现使用本地消息传递之间的代理并行计算的集合。相反的信念传播,我们的本地计划是保证收敛的平均值和所需的协方差量的全局解决方案,即使底层的因子图是loopy的情况下,同步更新,我们提供了收敛所需的迭代次数的限制。与信念传播相比,这种保证收敛的代价是额外的存储,计算和通信链路的情况下的循环;然而,我们展示了这些可以自动构建在飞行中只使用本地信息。

英文摘要:

Underlying many Bayesian inference techniques that seek to approximate the posterior as a Gaussian distribution is a fundamental linear algebra problem that must be solved for both the mean and key entries of the covariance. Even when the true posterior is not Gaussian (e.g., in the case of nonlinear measurement functions) we can use variational schemes that repeatedly solve this linear algebra problem at each iteration. In most cases, the question is not whether a solution to this problem exists, but rather how we can exploit problem-specific structure to find it efficiently. Our contribution is to clearly state the Fundamental Linear Algebra Problem of Gaussian Inference (FLAPOGI) and to provide a novel presentation (using Kronecker algebra) of the not-so-well-known result of Takahashi et al. (1973) that makes it possible to solve for key entries of the covariance matrix. We first provide a global solution and then a local version that can be implemented using local message passing amongst a collection of agents calculating in parallel. Contrary to belief propagation, our local scheme is guaranteed to converge in both the mean and desired covariance quantities to the global solution even when the underlying factor graph is loopy; in the case of synchronous updates, we provide a bound on the number of iterations required for convergence. Compared to belief propagation, this guaranteed convergence comes at the cost of additional storage, calculations, and communication links in the case of loops; however, we show how these can be automatically constructed on the fly using only local information.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑