Research About The Theories And Algorithms Of Transfer Learning

Serving as one of the most important research directions in machine learning, transfer learning has begun to catch more attention in recent years. In this paper, we will study some theories and algorithms about transfer learning and improve it by combining it with the adaboost algorithm, multi-view learning and multi-source learning.In general machine learning, according to the different study situation, requirement and hardness, we can divide transfer learning into several classes such as instance-transfer, parameter-transfer and relational-knowledge-transfer. In our paper, we will focus on the study of instance-transfer and parameter-transfer.Firstly, before studying the algorithms of transfer learning, we need to make a research about one famous ensemble learning algorithm, adaboost. Through some previous studies about this algorithm, we know the core of it is to generate a lot of weaker learners about one data set and combine them into one strong learner finally by a certain rule. In this paper, in order to promote the performance of the adaboost algorithm, we present the embedded multi-view adaboost algorithm by combining adaboost with multi-view learning. In the past studies, researchers always combine these two approaches linearly which can not exert the characteristics of them sufficiently. However, the purpose of the algorithm proposed by us is actually to overcome this problem by integrating them into a whole. In addition, we not only make an analysis about the embedded multi-view adaboost algorithm, but also do some experiments about it. According to the experimental results, we can see that the effectiveness of the embedded multi-view adaboost algorithm is better than the general adaboost algorithm.Instance-transfer is a fundamental and effective approach which intends to improve the learning effect of the target task by reusing the samples coming from the source task. Because of the simplicity and convenience of instance-transfer, this approach has been used in various fields. Moreover, due to the reason that we believe multi-view learning can take a great effect on transfer learning, we combine the embedded multi-view adaboost algorithm with transfer learning and propose a new algorithm named multi-view transfer learning with adaboost. This algorithm proves to be effective by some experiments on it. Parameter-transfer is another common approach of transfer learning, which parameterizes the information of the source task and transfer them to the target task. Now we intend to combine it with part-based model and propose one new algorithm named part-based transfer learning. Dissimilar to many traditional works, we consider how to transfer information form one task to another in the form of consequent parts. We regard all the tasks as a collection of several constituent parts and every task can be divided into several parts respectively. It means transfer learning between two tasks such as the source task and target task can be accomplished by sub-transfer learning between their parts. We believe this kind of way can exploit more information than before. The experimental results indicate that part-based transfer learning outperforms the general transfer learning.Though transfer learning is an effective and practical approach, it also contains some defects such as negative transfer. The appearance of negative transfer mostly depends on the difference between the source task and target task. For the purpose of avoiding it, we make use of multi-source learning. In stead of using only single source task, now we use multiple source tasks simultaneously to promote the probability of finding the one which is more similar to the target task. In our paper, multi-source learning will be used in both multi-view transfer learning with adaboost and part-based transfer learning. The experimental results of them validate the effectiveness of our proposed algorithms.