摘要
手写体数字识别网络的训练过程需耗费大量时间,训练时间的优化有着重要的意义。利用专家乘积系统是一种理想的解决方法。在专家乘积系统训练过程中,每个数字都将建立一个独立的专家模型,并分别使用各自的样本进行训练。待获得所有模型的概率分布特征后,再送入一个分类器网络进行混合训练。由于各数字模型是独立的,因此利用并行训练可大大减少系统的训练时间。专家乘积系统的识别效果非常理想,反映出专家乘积系统是一个高效的模型。
For recognizing handwritten digits , the training process of the network to need to waste a great deal of time. How to optimizing training time is importance. We can use the Products of Experts to solve it. During the training period, Every digit need to set up a expert model to have a training which only for samples of themselves so that the characteristic of probability distribution of different digital model could get from it. Probability distribution got from mixed training sample under different model will be sent to a classification network to be the training sample. Each digital model is independent so that trained in parallel could decrease system test time. Though the testing we could get pretty ideal recognizing result. This will turn out that Products of Experts is really a perfect model.
出处
《计算机仿真》
CSCD
2006年第5期197-199,214,共4页
Computer Simulation
关键词
专家乘积
玻耳兹曼机
分类网络
手写体识别
Products of experts
Bohzmann machine
Classification network
Handwriting recongnition