基于专家乘积的手写体数字识别系统-计算机应用专业论文.docxVIP

AbstractPassing Abstract Passing many years studies，the recognizing handwritten digits have obtains the good result and by used in many fields．Mainly method by used for recognizing of the neural network iS include SVM，feedback network etc．For acquire better recognizing result，more hide layers and more cell are use in the networks．This makes the whole network of training time increase consumedly．The method of mixture model can proceed together to train each expert system．But in common mixture model，it can not get well result when classifying high—dimensional data．The method Products of Experts work well to resolve thiS problem． ThiS paper mainly introduced recognizing handwritten digits to use the PoE．In the First，it was simple to introduce basiC info of the neural network．Then，the general method of the common model mixture i S di scussed． It point out the common mixture method iS inefficient in high——dimensional data．Pass the analysi S to the PoE theories，we find out standard of learning method from restricted boltzmann machine iS fully equal to the learning of PoE．We could use method of contrastive divergence to further improve Gibbs sampling then reduce training time． The improvement of recognizing of time is another topic will be expounded here．The big matrix operation always took up a great deal of time．Using decomposed Kronecker multiply，weight net can be divided into sum product of several small matrixes．We can use hardware to implement parallel computer in this way it could meet improve and shorten the recognizing of time． In this paper we will use PoE and the Decomposed Kronecker matrix to realize recognizing handwritten digital．Passing the analysis of result， we can prove it is an efficient method． Keywords： Products of Experts、recognizing handwritten digital、Decomposed Kronecker Matrix、Mixture Experts 第1I页 第一章绪论第1章绪论 第一章绪论 第1章绪论 现代计算机具有强大的计算和信息处理的能力，但是它在目标识别、环境感 知及在复杂条件下的决策能力远不如生物系统。目前，已有很多学科分别从不同 角度、以不同的途径试图研究和揭示这当中的奥秘，并且希望用计算机实现一个 具有感知、识别、理解、自学习和自适应能力的