Overview of transfer learning

Ministrante: Dr. Ricardo Prudêncio, Universidade Federal de Pernambuco


Transfer Learning is a paradigm in machine learning which aims to improve performance in a target learning domain by taking advantage of rich information available in a source domain. The literature of the area is huge and the techniques may vary a lot depending on the basic assumptions concerning both the target and the source domain. For instance, in some cases it is assumed that there is a small labeled training set in the target domain, while in other contexts no labeled data is available, thus requiring reusing source labeled instances to learn a supervised model. As another example, both target and source can have the same features (i.e., homogeneous transfer learning), while in heterogeneous transfer learning, one can reuse information in completely different feature spaces (for instance, reusing textual information to support image classification). In this short tutorial, we will discuss different approaches and types of transfer learning. Examples of each category will be presented as well as research challenges and directions. CV resumido: Ricardo Prudêncio is Associate Professor at Centro de Informática / Universidade Federal de Pernambuco. He received this Ph.D. at Computer Science from Universidade Federal de Pernambuco (2004) and participated as a postdoc at Bristol University (2015-2016). He has experience in Computer and Information Science, acting on the following subjects: machine learning, artificial neural networks, hybrid intelligent systems, social network analysis and text mining.