一种改进的BP神经网络在手写体识别上的应用

传统的浅层学习神经网络虽然结构简单,算法速度快,但错误率较高,且容易陷入局部最小。文中采用深度结构的深度置信网,优化基于传统BP神经网的初始值,以获得较好的检测结果,并利用Dropout技术改进BP网络隐层单元,获得较快的运算速度。实验证明,经过DBN和Dropout改善后的网络错误率有明显降低,并且算法实时性得到了一定改善。

Application of an Improved BP Neural Network in Handwriting Recognition

Traditional shadow learning has a simple structure and fast algorithm but a relatively high error rate,and is easy to step into the local minimization.This article discusses a deep structure call Deep Belief Network (DBN),which is used to optimize the initial value of the traditional BP to get a better testing result.And we use the dropout skill to modify the hidden unit of BP to get a fast training speed.The experiments in the article show that using DBN and dropout to modify the BP can decrease the error rate and improve the real timeness.