Application of Evolutionary Multiobjective Optimization in L1-Regularization of CNN
Application of Evolutionary Multiobjective Optimization in L1-Regularization of CNN
复制标题
DOI：
10.1109/cec.2019.8789899
复制
发表时间：
2019-06
期刊：
影响因子：
--
通讯作者：
Misaki Kitahashi;H. Handa
中科院分区：
文献类型：
--
作者：
Misaki Kitahashi;H. Handa
卷积神经网络（CNN）是一种多层神经网络，其在处理图像和自然语言方面具有上级结构。L1正则化可以用作防止CNN过拟合和提高泛化性能的手段。然而，很难设置适当的正则化项系数。因此，在本研究中，我们使用MOEA/D来确定超参数的值。MNIST上的实验结果表明了该方法的有效性。
A convolutional neural network (CNN) is a type of multilayered neural network, which has a structure superior in handling images and natural languages. L1-regularization may be used as a means to prevent overfitting of CNN and improve generalization performance. However, it is difficult to set an appropriate coefficient of the regularization term. Therefore, in this research, we use MOEA/D to determine the value of hyper parameter. Experimental results on MNIST show the effectiveness of the proposed approach.