一种基于修正动量的RBM算法

受限玻尔兹曼机（Restricted Boltzmann Machine，RBM）是一种随机网络、概率图模型，它是一种比较有效的的无监督学习模型.针对RBM梯度近似的一种计算方法对动量加速不敏感，以及识别效果不理想等问题，本文提出一种基于修正动量的RBM算法.该算法结合RBM梯度近似方法，通过修改隐单元偏置参数的更新方式，避免RBM模型中隐单元取值采用概率值时导致模型识别效果不理想、动量加速有限等问题.同时，在RBM预训练阶段采用快速上升的动量方式，以加速网络收敛；在微调阶段引入缓慢下降的动量项，以避免陷入局部最优点并提高识别效果.本文算法通过在MNIST手写数字体，Extended Yale B和CMU-PIE人脸数据库上的数值实验结果表明，提出的算法能够有效地提高计算效率和提高网络泛化能力.该算法不仅对RBM的应用领域扩展具有十分积极的实际意义，且为深度学习的应用方法提供一种新的研究思路和借鉴.

An Algorithm Based on Modified Momentum Using Restricted Boltzmann Machine

Restricted Boltzmann machine (RBM) is a stochastic neural network and probabilistic graphical model,which is one of the most effective models without supervision in deep learning.Focusing on the gradient approximation algorithm insensitivity to the momentum acceleration and recognition effectiveness in RBM,we propose the algorithm based on modified momentum using RBM.When the rule to update the hidden states adopts the probability value instead of sampling a binary value,this calculation method for the RBM gradient approximation leads to the undesirable recognition performance and limited momentum acceleration.Therefore,we modify the updating rule of the hidden bias to avoid these problems.Simultaneously,we use the rapidly ascending momentum method to improve the learning speed in the RBM pre-training phase.An improved slowly descending momentum method is also used in the fine-tuning stage to accurately find the best point,which is far from becoming trapped in poor local optima and improves the classification effect.Through the recognition experiments on MNIST dataset,Extended Yale B and CMU-PIE face dataset,the achieved results show that the proposed algorithm can enhance the computation efficiency and improve the generalization ability of networks.The algorithm not only extends the application fields of RBM,but also provides a new research idea and reference for the application method of deep learning.