基于相似度的神经网络多源迁移学习算法

为了解决迁移学习中的"负迁移"问题,提出了基于相似度的神经网络多源迁移学习算法。该算法是以经典的BP神经网络模型为基分类器,利用梯度下降法对各个源领域与目标域之间的相似度进行学习和优化,把各个源领域的网络权重参数信息按照与目标域之间的相似程度迁移到目标域中,提高机器学习算法在目标域的分类性能。在UCI数据的Letter-recognition数据集以及20Newsgroups文本数据集上进行实验。实验结果表明了MTL-SNN算法比传统的多源迁移学习算法以及BP神经网络算法在分类准确率上有所提升,因此MTL-SNN算法有效地解决了"负迁移"问题。

Research on Multi-source Transfer Learning Algorithm Based on Similarity Neural Network

In order to solve the problem of "negative transfer" in transfer learning, a similarity-based neural network multi-source transfer learning algorithm is proposed. The algorithm uses the gradient descent method to learn and optimize the similarity between each source and target domain, which is based on the classical BP neural network model. According to the degree of similarity with the target domain, the network weight parameter information of each source domain is migrated to the target domain, and the classification performance of the machine learning algorithm in the target domain is improved. The paper tests on the Letter-recognition data set of UCI data and the Newsgroup20 text data set. The experimental results show that the MTL-SNN algorithm has improved classification accuracy compared with the traditional multi-source migration learning algorithm and BP neural network algorithm. Therefore, the MTL-SNN algorithm effectively solves the "negative transfer" problem.