多视角和迁移学习识别方法和智能建模研究

发布时间：2018-06-13 14:51

本文选题：多视角学习 + 迁移学习　；参考：《江南大学》2015年博士论文

【摘要】：人工智能技术发展已近70年,在此期间各种智能化方法被提出用于解决各种实际应用问题,其中模糊识别技术及模糊智能系统建模技术在医疗、控制、经济等领域得到了广泛的关注及使用。然而,随着人们生活及科技水平的提升,越来越多的新应用场景被发现。在众多新应用场景中,多视角应用场景及迁移应用场景对人们的生活和生产有着广泛的影响,本课题将主要关注上述两个新兴的应用场景。在上述两新应用场景下,经研究我们发现一些经典的模糊识别技术及模糊智能系统建模技术的性能变得不再可靠,其通常面临以下几个挑战:1)在多视角应用场景下,由于经典的模糊识别技术及模糊智能系统建模技术均是在单视角学习场景下提出的,它们本身不具备多视角协同学习的能力,若执意选择这些经典方法进行多视角学习则只能在每个视角上独立学习,从而得到无法令人满意的结果;2)在迁移应用场景下,由于生产的保密性或一个产业本身就是一新兴产业以往并无数据积累,这使得经典的模糊识别技术及模糊智能系统建模技术在对该产业进行数据处理或学习建模时可用数据极少,从而导致经典方法的失效。为解决经典的模糊识别技术及模糊智能系统建模技术在面对上述新兴应用场景时所面临的问题,本课题将主要分两大部分分别在多视角及迁移场景下对模糊识别技术及模糊智能系统建模技术的相关方法做出适当的改进以期提高模型的性能,具体如下:(1)第一部分为第2至4章节,主要探讨了模糊识别技术及模糊智能系统建模技术在多视角领域的应用。首先,在第2章节针对一种已有的基于模糊聚类算法的多视角模型Co-FKM所存在的问题,基于Havrda-Charvat熵理论重新构造了新的多视角协同学习方法即异视角空间划分逼近准则,并针对多视角场景下各视角存在差异性的问题,基于香农熵理论提出了多视角自适应加权策略,进而提出了一种熵加权多视角协同划分模糊聚类算法EW-Co P-MVFCM。其次,基于第2章节关于多视角模糊聚类的认识,于第3章节基于先前的模糊聚类工作即GIFP-FCM算法,提出了一种具备多视角协同学习能力的基础算法Co-FCM,所提出的Co-FCM算法引入了4种不同的协同度量函数用于拓展方法的应用范围,进一步地,考虑到各视角存在的差异性而在基础的多视角模糊聚类Co-FCM算法上又加入了视角加权机制得到了最终的WV-Co-FCM算法,该算法较之Co-FCM算法不仅获得了最佳视角的辨识能力,同时还拥有了更佳的聚类性能。最后,于第4章节,探讨了模糊智能系统建模技术在多视角领域的应用,具体地先提出了一种基于大间隔分类机制的单视角模糊分类模型TSK-FC算法,并以此算法作为基础模型通过融入多视角协同学习机制得到了一种具备双视角协同学习能力的TSK模糊分类模型Two V-TSK-FC,该模型通过协同学习机制能够在建模过程中利用各视角独立信息的同时进一步利用视角间的关联信息增强算法的性能。(2)第二部分为第5至7章节,主要探讨了模糊识别技术及模糊智能系统建模技术在迁移学习领域的应用。首先,第5章节针对非充分数据集(数据贫乏)及噪声对最终聚类结果产生严重干扰的问题,依旧基于本章节所提及的一般化的增强模糊划分聚类算法(GIFP-FCM),以此算法为基础通过在该算法中融入具备聚类特性的迁移学习机制以使得GIFP-FCM算法获得迁移能力,最终得到迁移GIFP-FCM算法T-GIFP-FCM。其次,针对传统模糊系统建模方法在迁移场景下存在的问题,以广泛应用的TSK型模糊系统作为研究对象,探讨了具有迁移学习能力的模糊系统,即TSK型迁移学习模糊系统,所提的迁移学习TSK模糊系统不仅能充分利用当前场景的数据信息,还能有效地利用历史相关场景所积累得到的知识对当前源场景的建模过程进行辅助学习,从而提高模型的泛化性能。最后,第7章节进一步地针对第6章节提出的TSK型迁移模糊系统在模糊前件参数和后件参数迁移学习时所存在的一系列问题提出了相应的改进方案,具体地结合第5章节提出的迁移模糊聚类理论以及一种改进的迁移学习后件学习机制,提出了一种增强知识迁移的TSK迁移学习模糊系统,该方法的提出有效地将迁移聚类和迁移模糊系统建模相结合,使得模糊系统的建模过程更为智能且学习能力更为优秀,同时该方法的提出也为迁移学习在智能建模领域的发展提供一种新的研究思路。
[Abstract]:Artificial intelligence technology has been developed for nearly 70 years. During this period, various intelligent methods have been put forward to solve various practical application problems. Fuzzy recognition technology and fuzzy intelligent system modeling technology have been widely concerned and used in medical, control, economic and other fields. However, with the improvement of people's life and science and technology, more and more Many new application scenarios have been found. In many new application scenarios, multi view application scene and migration application scene have a wide impact on people's life and production. This topic will focus on the above two emerging application scenarios. Under the above two new application scenarios, we have found some classic fuzzy recognition techniques and models. The performance of paste intelligent system modeling technology is no longer reliable, and it usually faces the following challenges: 1) under the multi perspective application scene, because the classical fuzzy recognition technology and fuzzy intelligent system modeling technology are proposed in the single perspective learning scene, they do not have the ability of multi perspective collaborative learning, if they insist on choosing this Some classical methods for multi perspective learning can only be studied independently in every perspective, thus getting unsatisfactory results. 2) in the migration application scenario, the classic fuzzy recognition technology and fuzzy intelligent system modeling are made because of the secrecy of production or the industry itself is a new industry that has not accumulated data in the past. In order to solve the problems faced by the classical fuzzy recognition technology and the fuzzy intelligent system modeling technology in the face of the above emerging application scenarios, the subject will be divided into two parts in multi perspective and migration field respectively. The relevant methods of fuzzy recognition technology and fuzzy intelligent system modeling technology are improved to improve the performance of the model. The following is as follows: (1) the first part is chapter second to 4, mainly discusses the application of fuzzy recognition technology and fuzzy intelligent system modeling technology in the field of multi angle. There are some problems in the multi view model Co-FKM based on fuzzy clustering algorithm. Based on the Havrda-Charvat entropy theory, a new multi perspective cooperative learning method is rebuilt, that is, the different angle of view spatial partition approximation criterion. In view of the difference of various perspectives in multi view scenes, a multi angle adaptive addition based on Shannon entropy theory is proposed. Right strategy, and then proposed an entropy weighted multi view cooperative division fuzzy clustering algorithm EW-Co P-MVFCM. next, based on the second chapter on the understanding of multi perspective fuzzy clustering, the third chapter based on the previous fuzzy clustering work, GIFP-FCM algorithm, proposed a multi perspective collaborative learning ability of the basic algorithm Co-FCM, proposed. The Co-FCM algorithm introduces 4 different cooperative metric functions to extend the application scope of the method. Further, considering the differences in various perspectives, the final WV-Co-FCM algorithm is obtained by adding a visual angle weighting mechanism to the basic multi perspective fuzzy clustering Co-FCM algorithm. The algorithm is not only the best than the Co-FCM algorithm. At the same time, it also has better clustering performance. Finally, in the fourth chapter, the application of fuzzy intelligent system modeling technology in multi field of view is discussed. A single view fuzzy classification model TSK-FC algorithm based on large interval classification mechanism is put forward, and the algorithm is used as the basic model to integrate multi view. The angle cooperative learning mechanism has obtained a TSK fuzzy classification model, Two V-TSK-FC, which has the ability of dual perspective collaborative learning. Through collaborative learning mechanism, the model can make use of the independent information from various perspectives in the process of modeling and further enhance the performance of the algorithm. (2) the second part is the fifth to 7 chapters. The application of fuzzy recognition technology and fuzzy intelligent system modeling technology in the field of migration learning is discussed. First, the fifth chapter is still based on the general enhanced fuzzy partition clustering algorithm (GIFP-FCM), which is still based on this chapter, aiming at the serious interference of the incomplete data set (data poor) and noise to the final clustering results. By integrating the migration learning mechanism with clustering characteristics in the algorithm, the migration ability of GIFP-FCM algorithm is obtained, and then the migration GIFP-FCM algorithm T-GIFP-FCM. is finally obtained. In view of the problems existing in the traditional fuzzy system modeling method in the migration scene, the TSK type fuzzy system is widely used as the research object. The fuzzy system with migration learning ability, that is, TSK type migration learning fuzzy system, the proposed migration learning TSK fuzzy system can not only make full use of the data information of the current scene, but also effectively use the knowledge accumulated in the history related scene to study the modeling process of the current source scene, thus improving the model. Finally, the seventh chapter puts forward the corresponding improvement scheme for the TSK type migration fuzzy system proposed by the sixth chapter in the learning of the fuzzy precursor parameters and the migration of the post parameters, and specifically combines the migration fuzzy clustering theory with the fifth chapter and an improved migration learning. A TSK migration learning fuzzy system is proposed to enhance knowledge migration. The proposed method effectively combines migration clustering and migration fuzzy system modeling, making the modeling process of the fuzzy system more intelligent and better learning ability, and the method is also proposed for the migration learning in the field of intelligent modeling. Development provides a new way of thinking.
【学位授予单位】：江南大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP18;TP311.13

【相似文献】