基于类别条件的受限玻尔兹曼机改进设计OACSTPCD
Improvement of RBM Based on Label Condition
针对受限玻尔兹曼机(RBM )在进行无监督训练时易出现特征同质化导致泛化能力较差的问题,设计了将类别条件引入 RBM 训练中,从而提出了基于类别条件的 RBM (lCRBM )。针对 RBM 的训练,将类别信息作为模型隐单元训练条件,参与到隐单元后验激活概率计算中;并将该模型作为深度玻尔兹曼机(DBM )的底层结构,应用于深度学习中。通过手写数字识别集合测试,该模型在训练速度和特征提取有效性上均有较大改善,并且能够提高深度模型的特征学习能力。
In order to overcome the poor generalization resulted by characteristic homogeneity for unsupervised training of restricted boltzmann machine ,the label condition is introduced to the training of RBM to design a new model named label condition RBM .Training for RBM ,the label information is treated as training condition for RBM and involved in the calcula-tion of hidden units’ posterior probability .The model is applied in deep learning which designed as the…查看全部>>
贺鹏程
海军装备部驻重庆地区军事代表局 重庆 400042
信息技术与安全科学
受限玻尔兹曼机深度学习监督学习对比散度
restricted Boltzmann machinedeep learningsupervised learningcontrastive divergence
《计算机与数字工程》 2016 (8)
1436-1438,1547,4
评论