摘要
传统无监督迁移学习[1]仅考虑训练集来源于一个域,现实情景中,往往会有多个源域与目标域存在相似性。现有多源无监督学习算法只单一考虑源域与目标域之间的边缘分布或条件分布,或简单对半叠加。一般来说不同的源域和目标域,其边缘分布与条件分布对于良好的迁移性能的贡献是不一样的,并且对抗式迁移学习相比非对抗有较好的效果且模型更具有鲁棒性。为了解决上述问题,提出了一种基于平衡分布自适应的多源深度对抗迁移学习算法(MATLBDA)。MATLBDA算法通过调节两大分布的权重,来解决多源迁移学习中分布自适应问题。相关实验表明,该方法在Office-31、DomainNet等主流图像分类数据集上有着良好的效果。
Traditional unsupervised transfer learning only considers that the training set comes from one domain.In real situations,there are often multiple source domains that are similar to the target domain.Existing multi-source unsupervised learning algorithms only consider the edge distribution or conditional distribution between the source domain and the target domain,or simply superposition in half.In general different source domain and target domain,its marginal distribution and the conditional distribution for the contribution of good migration performance is not the same,and migration of adversary learning than confrontation has good effect and the model is more robust.In order to solve the above problem,a multil-source deep adversarial transfer learning algorithm based on balanced distribution adaption is proposed.MATLBDA algorithm solves the problem of distribution adaptation in multi-source transfer learning by adjusting the weights of two large distributions.Experiments show that the method has good performance on mainstream image classification datasets such as Office-31 and DomainNet.
作者
王业胜
杜景林
WANG Yeshengg;DU Jinglin(School of Artifical Intelligence(School of Future Technology),Nanjing University of Information Science&Technology,Nanjing 210044)
出处
《计算机与数字工程》
2025年第6期1728-1733,共6页
Computer & Digital Engineering
关键词
迁移学习
多源域
注意力机制
对抗学习
transfer learning
multi-source domain
attention mechanism
adversarial learning
