迁移学习

一什么是迁移学习：

        给出源领域的数据Ds和任务Ts，目标领域的数据Dt和任务Tt，迁移学习旨在使用源领域Ds和Ts中的知识去改进对于目标领域的预测函数ƒt()，其中Ds≠Dt或者Ts≠Tt

二、为什么需要迁移学习：

      a 数据的标签很难获取

      b从头建立模型是复杂和耗时的

 

三、迁移学习解决了常规学习解决不了的问题：

• Why to transfer

     1. Source and target domains do not need to be in the same distributions.

         可解决数据分布不同的问题

    2.Doesn’t need a large amount of data. Less training samples, even none.

        可解决小数据量问题

     3. Building every model from scratch is time-consuming and expensive.

          运用已有知识来学习

四、运用迁移学习要注意的：

• What to transfer：which part of knowledge can be transferred across

                                   domains or tasks

       哪部分知识可以迁移到task上去

• when to transfer： in which situations, transferring skills should be   

                                  done. when the source domain and target domain 

                                  are not related to each other, brute-force transfer 

                                  may be unsuccessful

        需要结合具体场景，源域和目标域没有相关性就不能强行迁移学习

• How to transfer： source and target domains be related to each 

                                 other. However, how to avoid negative transfer is

                                an important open issue that is attracting more and

                                more attention in the future

     主要避免negative transfer

negative transfer： Domains are too dissimilar

                                    Conditional Kolmogorov complexity is not related

                                    Tasks are not well-related

五、迁移方法的分类