图 书 馆
应用于电磁场算法的三级加速及其基于外推的变型Triple jump acceleration for the EM algorithm and its extrapolation-based variants |
|
| 课程网址: | http://videolectures.net/mlss06tw_huang_tjaai/ |
| 主讲教师: | Han-Shen Huang |
| 开课单位: | 中央研究院 |
| 开课时间: | 2007-04-25 |
| 课程语种: | 英语 |
| 中文简介: |  |
| 课程简介: | The Aitken's acceleration is one of the most commonly used method to speed up the fixed-point iteration computation, including the EM algorithm. However, it requires to compute or approximate the Jacobian of the EM mapping matrix, which can be intractable for complex models. We will present our current research topic, the triple-jump acceleration, to accelerate the EM and some of its extrapolation-based variants by approximating their Jacobians. One advantage of the triple jump framework is that we can directly use the EM and its extrapolation-based variants as black boxes and achieve acceleration easily. We can update parameter vectors globally with the same approximated Jacobian, or locally based on each decomposable component with different Jacobians. Experimental results show that the triple jump methods consistently accelerate EM, parameterized EM (pEM) and adaptive EM (aEM) for a variety of probabilistic models. |
| 关 键 词: | 迭代计算; 电磁算法; 雅可比矩阵; 概率模型 |
| 课程来源: | 视频讲座网 |
| 最后编审: | 2020-06-26：cxin |
| 阅读次数: | 33 |