原标题：Transitive Transfer Learning
作者：Ben Tan, Yangqiu Song, Erheng Zhong, QiangYang
中文摘要：
迁移学习是利用源领域的知识来提高目标领域的学习能力的一种学习方法，在各种应用中被证明是有效的。迁移学习的一个主要限制是源域和目标域应该是直接相关的。如果两个领域之间的重叠很少，那么在这些领域之间进行知识转移将是不有效的。受人类传递推理和学习能力的启发，我们研究了一个新的学习问题:传递迁移学习(transitive Transfer learning，缩写TTL)。TTL旨在打破大的领域距离和转移知识，即使源和目标领域直接共享很少的因素。例如，当源域和目标域分别是文本和图像时，TTL可以使用一些带注释的图像作为中间域来桥接它们。为了解决TTL问题，我们提出了一个框架，首先选择一个或多个域作为源域和目标域之间的桥梁，实现迁移学习，然后通过这个桥梁进行知识迁移。广泛的经验证据表明，该框架在几个分类数据集上产生了最先进的分类精度。
英文摘要：
Transfer learning, which leveragesknowledge from source domains to enhance learning ability in a target domain,has been proven effective in various applications. A major limitation oftransfer learning is that the source and target domains should be directlyrelated. If there is little overlap between the two domains, performingknowledge transfer between these domains will not be effective. Inspired byhuman transitive inference and learning ability, whereby two seeminglyunrelated concepts can be connected by series of intermediate bridges usingauxiliary concepts, in this paper we study a novel learning problem: TransitiveTransfer Learning (abbreviated to TTL). TTL is aimed at breaking the largedomain distances and transferring knowledge even when the source and targetdomains share few factors directly. For example, when the source and targetdomains are text and images respectively, TTL can use some annotated images asthe intermediate domain to bridge them. To solve the TTL problem, we propose aframework wherein we first select one or more domains to act as a bridgebetween the source and target domains to enable transfer learning, and thenperform the transferring of knowledge via this bridge. Extensive empirical evidenceshows that the framework yields state-of-the-art classification accuracies onseveral classification data sets.
原文链接：https://dl.acm.org/doi/pdf/10.1145/2783258.2783295
原标题：Distant Domain Transfer Learning
作者：Ben Tan , Yu Zhang , Sinno JialinPan , Qiang Yang
中文摘要：
本文研究了一个新的迁移学习问题——远域迁移学习(DDTL)。与现有的迁移学习问题假设源域和目标域之间存在密切关系不同，在DDTL问题中，目标域可以与源域完全不同。例如，源域对人脸图像进行分类，而目标域对平面图像进行区分。受人类认知过程的启发，通过逐步学习中间概念，两个看似不相关的概念可以连接起来，我们提出了一种选择性学习算法(Selective learning Algorithm, SLA)来解决DDTL问题，该算法使用有监督自动编码器或有监督卷积自动编码器作为基础模型来处理不同类型的输入。从直观上看，SLA算法从中间领域逐渐选择有用的未标记数据作为桥梁，以打破两个遥远领域之间巨大的分布差距来传递知识。对图像分类问题的实证研究证明了该算法的有效性，在一些任务上，与“非转移”方法相比，分类准确率提高了17%。
英文摘要：
In this paper, we study a novel transferlearning problem termed Distant Domain Transfer Learning (DDTL). Different fromexisting transfer learning problems which assume that there is a close relationbetween the source domain and the target domain, in the DDTL problem, thetarget domain can be totally different from the source domain. For example, thesource domain classifies face images but the target domain distinguishes planeimages. Inspired by the cognitive process of human where two seemingly unrelatedconcepts can be connected by learning intermediate concepts gradually, wepropose a Selective Learning Algorithm (SLA) to solve the DDTL problem withsupervised autoencoder or supervised convolutional autoencoder as a base modelfor handling different types of inputs. Intuitively, the SLA algorithm selectsusefully unlabeled data gradually from intermediate domains as a bridge tobreak the large distribution gap for transferring knowledge between two distantdomains. Empirical studies on image classification problems demonstrate theeffectiveness of the proposed algorithm, and on some tasks the improvement interms of the classification accuracy is up to 17% over “non-transfer” methods.
原文链接https://www3.ntu.edu.sg/home/sinnopan/publications/[AAAI17]Distant%20Domain%20Transfer%20Learning.pdf
文献总结：
