摘要
传统机器学习方法在识别领域或建模分析时其性能往往依赖于数据的丰富程度,训练数据或待分析数据量的多少决定最终模型的性能,通常数据量越多所构模型的性能越佳。但对于一些特殊领域或新兴领域,初期所采集的数据量常显不足,此时传统机器学习方法不再有效。针对此问题,近年来一种新的学习模式迁移学习方法被提出。该方法可通过对历史相关数据的总结获取历史知识,利用知识和当前的数据进行识别或建模从而得到更可靠的模型。该学习过程与人类的认知过程类似,其开创了一种新的学习模式。介绍和总结近年来在迁移学习框架下提升传统模糊识别方法(主要模糊聚类技术)和模糊智能建模方法(主要模糊系统建模技术)性能所做的一系列工作及尚存在的一些问题,以期为相关领域的研究人员提供有价值的参考。
In pattern recognition or modeling,the training and testing data using the traditional machine learning methods must satisfy the same distribution. In addition,the number of training data and/or the data to be analyzed also determines the performance of the final models. Typically,the more the amount number of data,the better the performance of the final models. However,data are often inadequate or insufficient in some special scenarios. In these cases,the conventional machine learning methods are invalid. To solve this problem,a new learning strategy called the transfer learning method has been proposed in recent years. This strategy can be used to acquire knowledge by mining the historical data,and then using the obtained historical knowledge and the current data builds a more reliable model.The learning process is similar to human cognitive processes which created a new learning model. This paper introduces and summarizes a series of our works in recent years,which are mainly about how to enhance the performance of the traditional fuzzy recognition method( the fuzzy clustering technique) and the fuzzy intelligent modeling methods( fuzzy systems modeling techniques) in the transfer learning scenarios. This work will provide a valuable reference for researchers in related fields.
出处
《江南大学学报(自然科学版)》
CAS
2015年第4期489-497,共9页
Joural of Jiangnan University (Natural Science Edition)
关键词
迁移学习
历史知识
模糊识别方法
模糊智能建模方法
transfer learning,history knowledge,fuzzy recognition method,fuzzy intelligent modeling method