基于牛顿方法的玻尔兹曼机训练

Boltzmann Machine Learning Based on Newton-type Method

使用二阶牛顿方法训练玻尔兹曼机,通过分析和实验验证发现:与随机梯度下降优化方法相比,牛顿方法训练的速度更快,并且在样本较小时能够获得更好的训练效果.但是,牛顿方法存在一定的问题,即在搜索到最优点之后训练结果不收敛.因此,必须进一步改进牛顿方法,使得训练过程趋于收敛,或者找到合适的跳出准则来终止训练.

The experimental verification and analysis of the training of the Boltzmann machine with the second-order Newton method found that the speed of training with Newton method is faster than that of the SGD method,and the training result can be better when the sample is small.However,the training result does not come down after reaching the best point.It is necessary to further improve the Newton method to address the problem or to find a suitable jumping-out criteria to terminate the training.