一种具有迁移学习能力的RBF-NN算法及其应用

A RBF-NN algorithm with transfer learning ability and its application

经典的径向基人工神经网络学习能逼近任意函数,因而应用广泛。但其存在的一个重要缺陷是,在已标签样本过少、不能反映数据集整体分布情况下,容易产生过拟合现象,从而导致泛化性能严重下降。针对上述问题,探讨具有迁移学习能力的径向基人工神经网络学习算法,该算法在引入不敏感损失函数和结构风险项的同时,学习源领域径向基函数的中心向量及核宽和源领域模型参数,通过充分学习历史源领域知识来弥补当前领域因已标签样本少而导致泛化能力下降的不足。将该算法应用于人造数据集和真实发酵数据集进行验证,和传统的RBF神经网络算法相比,所提算法在已标签样本少而存在数据缺失的场景下,具有更好的适应性。

The classical radial basis function neural network(RBF-NN)is widely used as it can approximate any function.However,one of its main defects is that overfitting is likely to occur when there are too few labeled samples to reflect the overall distribution of datasets;this leads to a serious decline in its generalization ability.To solve the above problem,an artificial RBF-NN learning algorithm with transfer learning ability is discussed.The algorithm introduces the-insensitive loss function and the structural risk term and also learns the center vector and kernel width of the radial basis function as well as the parameters of the source domain model.The algorithm fully learns the knowledge in the historical source domain to compensate for its decline in generalization ability caused by the lack of labeled samples in the current field.To verify the algorithm,it is applied to an artificial dataset and real fermentation dataset.Compared with the traditional RBF-NN algorithm,the proposed algorithm has a better adaptability as regards less labeled samples and missing data.