基于数据集特征的卷积网络代价函数优化方法

Convolution neural network cost function optimization method based on Dataset features

近年来,为了减少卷积神经网络计算量和提高分类准确率,研究者们开发了各种各样的建模技术和优化方法.卷积神经网络中代价函数的作用是对比预测值和样本标签,评估预测准确程度.由于数据集中各类样本的分布也会一定程度上影响预测结果,提出一种基于数据集统计特征的代价函数优化方法,将数据集的统计特征以权重的形式叠加到交叉熵代价函数中.经过Mnist和Cifar数据集实验验证,表明该损失函数在卷积网络中收敛速度和准确率均有一定的优势.

In recent years,in order to reduce the convolution neural network computation and improve classification accuracy,the researchers developed a variety of modeling techniques and optimization method.The function of loss function is to evaluate the accuracy of prediction by comparing the predicted value and the sample label.Because all kinds of distribution of the sample in the data set may affect the forecasting result,put forward an optimization method of the cost function based on the characteristics of the data sets,then the statistical characteristic of data sets is reflected into the loss function as weights.With Mnist and Cifar data experiment verification,the results show that this new loss function has some advantages in converging rate and accuracy.