Fast /spl alpha/-weighted EM learning for neural networks of module mixtures
Fast /spl alpha/-weighted EM learning for neural networks of module mixtures
复制标题
用于模块混合神经网络的快速 /spl alpha/加权 EM 学习
DOI：
10.1109/ijcnn.1998.687221
复制
发表时间：
1998
期刊：
影响因子：
--
通讯作者：
T. Ikeda
中科院分区：
文献类型：
--
作者：
Y. Matsuyama;S. Furukawa;N. Takeda;T. Ikeda
一类扩展对数用于导出 /spl alpha/-加权 EM（/spl alpha/-加权期望最大化）算法。这些扩展的 EM 算法（WEM、/spl alpha/-EM）预计在速度上优于传统（对数）EM 算法。传统方法属于新 WEM 的特例。本文首先给出一般性的理论讨论。然后，针对专家混合神经网络的情况给出了明确的证据，表明其收敛速度比普通 EM 方法更快。这个过程需要三个步骤。第一步是展示具体算法。然后，从理论上检查收敛性。第三，尝试通过专家混合学习实验来证明WEM的优越性。除了监督学习之外，还检查了高斯混合的无监督情况。再次观察到 WEM 更快的收敛。
A class of extended logarithms is used to derive /spl alpha/-weighted EM (/spl alpha/-weighted expectation-maximization) algorithms. These extended EM algorithms (WEMs, /spl alpha/-EMs) have been anticipated to outperform the traditional (logarithmic) EM algorithm on speed. The traditional approach falls into a special case of the new WEM. In this paper, general theoretical discussions are given first. Then, clear-cut evidence that shows faster convergence than the ordinary EM approach are given for the case of mixture-of-expert neural networks. This process takes three steps. The first step is to show specific algorithms. Then, the convergence is theoretically checked. Thirdly, experiments on the mixture-of-expert learning are tried to show the superiority of the WEM. Besides the supervised learning, the unsupervised case for a Gaussian mixture is also examined. Faster convergence of the WEM is observed again.