Instance-based And Feature-based Transfer Learning

Traditional machine learning techniques make a basic assumption that the training andtest data should be under the same distributions. However, in many cases, this identical-distribution assumption does not hold. The violation of the assumption might happen whenthe training data are out of date, but new data are expensive to label. This leaves plenty oflabeled examples that are under a similar but different distribution, which is a waste throwaway entirely. In this situation, transfer learning becomes important to take the role of lever-aging these existing data knowledge.Transfer learning aims at using learned knowledge from one context to benefit fur-ther learning tasks in other contexts. Thus, transfer learning does not make the identical-distribution assumption as tractional machine learning algorithms. In this thesis, we broadlyreview the whole field of transfer learning, and then introduce our recent work on transferlearning accordingly. Our work can be divided into two parts: instance-based transfer learn-ing, and feature-based transfer learning. We will show that instance-based transfer learninghas better strength in knowledge transferring, while feature-based transfer learning is withmore generality.We present two transfer learning algorithms based on boosting technique and featuretranslation respectively. These two algorithms corresponds to instance-based and feature-based transfer learning. Our extensive experiments show that our algorithms can greatlyimprove several state-of-the-art algorithms in the situation of transfer learning, includingnear transfer and far transfer.