基于一致性正则化与熵最小化的半监督学习算法

Semi-supervised Learning Algorithm Based on the Consistency Regularization and Entropy Minimization

在一致性正则化与熵最小化的基础上提出一种新的半监督学习算法Mean Mixup,集成数据的互补信息,然后使用熵最小化给未标记数据生成可靠的伪标签,在一致性正则化下进一步优化模型分类结果。在常用数据集SVHN和CIFAR10上对Mean Mixup算法进行了评估,实验结果表明,所提出的方法在分类准确率上优于一些已有的半监督学习算法。

A new semi-supervised learning algorithm Mean Mixup was proposed,which was based on the consistency regularization and entropy minimization.Complementary information of data was fused,and then entropy minimization was used to generate reliable pseudo labels for unlabeled data.The classification results of the model were further optimized by consistency regularization.The Mean Mixup algorithm was evaluated on the commonly used datasets SVHN and CIFAR10.The experimental results showed that the proposed method was superior to some existing semi-supervised algorithms in classification accuracy.