基于迁移学习的径向基函数神经网络学习

Transfer learning based radius-basis-function neural-network modeling

现实场景中存在很多小样本量数据集而且多有失真,传统神经网络在处理这类数据时泛化能力较差,不能达到预测数据或分类的目的。迁移学习可通过学习数据集A有用的知识对与其相关但不同正态分布的小样本数据集B进行辅助学习,因此提出了具有迁移学习能力的神经网络,以实现更好的分类或逼近效果。以基于ε-不敏感准则和结构风险的径向基神经网络(RBF)为基础构造了迁移径向基神经网络(T-RBF-NN)。通过加噪音数据集实验以及真实数据集实验验证加入迁移学习的神经网络在小样本情况下比传统神经网络具有更好的泛化性和鲁棒性。

Real datasets always have small samples and are noised which can＇t be classified or have better regression result by traditional neural network. This paper proposes transfer learning neural network because it can assist small sample dataset B through the useful knowledge of dataset A to have better regression or classification result while A is similar with B and have different normal distribution. A Transfer learning Radius-Basis-Function Neural-Network（T-RBF-NN）modeling method is proposed based on ε-insensitive criterion and structure risk based RBF-NN modeling method. The experiments on data sets with noise and real scenarios datasets show that T-RBF-NN extends the applicability of the traditional neural-network modeling.