Abstract: Current transfer learning model studies the source data for future target inferences within a major view that the whole source data should be used to explore the shared knowledge structure. However, due to the limited availability of human ranked source domain, this assumption may not hold due to the fact that not all prior knowledge in the source domain is correlative to the target domain in most real-world applications. This paper proposes a general framework referred to discriminative knowledge-leverage (KL) based on generalized empirical risk minimization (GERM) transfer learning, where the empirical risk minimization (ERM) principle is generalized to the transfer learning setting. Additionally, this paper theoretically shows the upper bound of generalized ERM (GERM) for the practical discriminative transfer learning. The proposed method can alleviate negative transfer by automatically discovering useful objects from source domain. Extensive experiments verify that the proposed method can significantly outperform the state-of-the-art transfer learning methods on several artificial/public datasets.

Key words: transfer learning, empirical risk minimization (ERM), generalized empirical risk minimization (GERM), discriminative knowledge-leverage, negative transfer

摘要： 目前的迁移学习模型旨在利用事先准备好的源域数据为目标域学习提供辅助知识，即从源域抽象出与目标域共享的知识结构时，使用所有的源域数据。然而，由于人力资源的限制，收集真实场景下整体与目标域相关的源域数据并不现实。提出了一种泛化的经验风险最小化选择性知识利用模型，并给出了该模型的理论风险上界。所提模型能够自动筛选出与目标域相关的源域数据子集，解决了源域只有部分知识可用的问题，进而避免了在真实场景下使用整个源域数据集带来的负迁移效应。在模拟数据集和真实数据集上进行了仿真实验，结果显示所提算法较之传统迁移学习算法性能更佳。

关键词: 迁移学习, 经验风险最小化（ERM）, 泛化的经验风险最小化（GERM）, 区别性知识利用, 负迁移