基于深度学习网络的学前教育质量评价研究

Research on preschool education quality evaluation based on deep learning network

为解决浅层学习网络对复杂函数的表达能力与泛化能力受到限制的问题,提升学前教育质量评价精度,提出基于深度学习网络的学前教育质量评价方法。从学前教育环境创设、学前教育保教质量和幼儿发展三方面出发,构建包含三个一级指标和九个二级指标的学前教育质量评价指标体系。将评价指标体系中二级指标作为深度学习网络输入,利用无监督的预训练模型优化深度学习网络各层权值,基于由下至上的非监督学习过程确定受限玻尔兹曼机内各层的条件概率分布与联合概率分布,输出层依照输入的DMOS值整体,构建抽象一级指标与DMOS值间的回归模型,依照回归模型预测获取学前教育质量的客观评价结果。测试结果显示该方法评价结果与主观评价结果之间的线性相关系数与等级相关系数更接近于1。

Since the expression ability and generalization ability of shallow learning network for complex functions arerestricted,and the evaluation accuracy of the preschool education is unsatisfied,a preschool education quality evaluationmethod based on deep learning network is proposed. The quality evaluation index system of preschool education including 3 firstgrade indexes and 9 second grade indexes is constructed,in which the environment creation of preschool education,the qualityof preschool education and the early children development are considered. The second grade indexes in the evaluation indexsystem are taken as the input of deep learning network. The unsupervised pretraining model is used to optimize the weights ofeach layer of the deep learning network. The conditional probability distribution and joint probability distribution of each layer inthe restricted Boltzmann machine are determined on the basis of the unsupervised learning process from bottom to top. In theoutput layer,the parameters of each layer are wholly optimized according to the input DMOS(datamining optimization system)value to construct the regression model between the abstract first grade index and the DMOS value. According to this regressionmodel,the objective evaluation results of preschool education quality are predicted and obtained. The test results show that thelinear correlation coefficient and grade correlation coefficient between the evaluation results of the proposed method and thesubjective evaluation results are closer to 1.