摘要
针对深度信念网络（Deep Belief Networks，DBN）微调阶段过度依赖梯度而导致很难获取最优输出权值的问题，提出一种基于偏最小二乘回归（Partial Least Square Regression，PLSR）的DBN输出权值确定新方法。通过PLSR和DBN相结合，实现对DBN最后一个隐含层状态进行主成分提取，在最后一个隐含层与输出层之间建立PLSR模型。以更精确的特征来确定更好的输出权值。在一系列标准数据集上的实验结果表明，该方法能够获取更好的DBN输出权值，从而提高DBN的性能。
The fine-tuning of deep belief network （DBN） is too dependent on gradient descent and difficult to obtain the optimal output weights, a new method to determine the output weights of DBN based on partial least square regression （PLSR） is proposed to solve this problem. The principal component in the last hidden layer state of DBN is extracted by combining PLSR and DBN, PLSR model is built between the last hidden layer of DBN and the output layer, so that better output weights are determined with more accurate features. The experimental results on a series of standard datasets show that the method proposed can obtain better DBN output weights, and the performance of DBN is correspondingly improved.
作者
索明何
程乐
SUO Ming-he;CHENG Le(Electronic Engineering Research Center, Huai＇an College of Information Technology, Huai＇an 223003, Chin)
出处
《控制工程》
CSCD
北大核心
2018年第4期668-676,共9页
Control Engineering of China
基金
江苏省高校自然科学基金项目(16KJB520049)
关键词
深度信念网络
输出权值
偏最小二乘回归
主成分提取
Deep belief network
output weightsi partial least square regression
principal componentextraction
