贝叶斯非参的流式分布式变分推断
机器学习
2015-11-02 v1 机器学习
摘要
本文提出了一种为贝叶斯非参(BNP)模型创建流式、分布式推断算法的方法论。在所提出的框架中，处理节点接收数据小批次(minibatch)序列，为每个计算变分后验，并对中心模型进行异步流式更新。与先前的算法相比，所提框架是真正的流式、分布式、异步、无学习率且无条件截断的。开发该框架的关键挑战源于 BNP 模型不对其分量施加内在排序这一事实，即在每次更新前寻找小批次与中心 BNP 后验分量之间的对应关系。为此，本文构建了关于分量对应关系的组合优化问题，并提供了高效的求解技术。最后将该方法应用于 DP 混合模型，实验结果展示了其实际可扩展性与性能。
引用
@article{arxiv.1510.09161,
title = {Streaming, Distributed Variational Inference for Bayesian Nonparametrics},
author = {Trevor Campbell and Julian Straub and John W. Fisher and Jonathan P. How},
journal= {arXiv preprint arXiv:1510.09161},
year = {2015}
}
备注
This paper was presented at NIPS 2015. Please use the following BibTeX citation: @inproceedings{Campbell15_NIPS, Author = {Trevor Campbell and Julian Straub and John W. {Fisher III} and Jonathan P. How}, Title = {Streaming, Distributed Variational Inference for Bayesian Nonparametrics}, Booktitle = {Advances in Neural Information Processing Systems (NIPS)}, Year = {2015}}