本文将具有增益向量自适应的随机梯度优化方法——随机Meta-Des.(SMD)应用于条件随机场(CRF)的训练。

We apply Stochastic Meta-Descent (SMD), astochastic gradient optimization method with gain vector adaptation, to thetraining of Conditional Random Fields (CRFs).

在几组大数据集合上，优化器的结果收敛到相同质量的求解，其速度比有限记忆BFGS快一个数量级，这是迄今报告的领先方法。

On several large data sets, the resultingoptimizer converges to the same quality of solution over an order of magnitudefaster than limited-memory BFGS, the leading method reported to date.

我们报告了精确和不精确推断技术的研究结果。

We report results for both exact andinexact inference techniques.

最近，条件随机场（CRF）在机器学习领域得到了普及。

Conditional Random Fields (CRFs) haverecently gained popularity in the machine learning community.

目前对CRFs的训练方法包括广义迭代缩放（GIS）、共轭梯度（CG）和有限记忆BFGS。

Current training methods for CRFs includegeneralized iterative scaling (GIS), conjugate gradient (CG), andlimited-memory BFGS.

这些都是只用于批处理的算法，不能很好地工作于在线条件下，并且需要许多遍历训练数据才能收敛。

These are all batch-only algorithms that donot work well in an online setting, and require many passes through thetraining data to converge.

这限制了CRFs的可扩展性和适用性。

This currently limits the scalability andapplicability of CRFs to large real-world problems.

部分源码地址：

https://www.cs.ubc.ca/~murphyk/Software/CRF/crf2D_usage.html

下载英文原文及源码地址：

http://page2.dfpan.com/fs/7lcj7221b291b6800c5/

更多精彩文章请关注微信号：