可扩展迁移学习中跨媒体复杂问题自动映射研究

Traditional machine learning methods assume the training sets and test sets share the same distributions, however, in many applications, data distributions are dynamically changeable. In contrast, transfer learning considers that the related tasks and target tasks have different distributions. Then, methods of this kind can deal with dynamic data, which is useful to design better intelligent machine learning methods than traditional ones. In the research of transfer learning, there still exists some crucial problems: 1) how to construct the scalable models among complex tasks, 2) how to choose the number of dimensionality in feature spaces, and 3) how to achieve automatical mapping among complex tasks. To solve the problems described above, we employ and reconstruct the following methods: Hierarchical Bayesian, deep leaning and incremental learning methods are used to build the representation models of abstract knowledge between different tasks; Informtion Entropy, Minimum Description Length (MDL) and Affinity Propagation (AP) are used to decide the number of feature dimensions; Multivariate information bottleneck (MIB), and matrix theorem based methods are used to obtain automatic mapping. In practical applications, this project uses different transfer learning models to study the pattern recognition and automatic mutual annotation in cross-media, and try to find the properties and mathematical rules existing in transfer learning for solving appropriate problems. This research, on the one hand, can benefit for designing more intelligent machine learning algorithms, and providing novel ideas and approaches to the development of artificial intelligence, on the other hand, by using transfer learning methods, can establish stable foundation for analyzing large scale data, recognizing the patterns in multimedia and building dynamic fast decision systems.

迁移学习打破了传统机器学习方法中训练集(相关任务)和测试集(目标任务)服从同分布的假设，能够处理分布发生变化的数据，有助于设计更加智能化的机器学习方法。本项目针对迁移学习中存在的关键问题：复杂任务间可扩展模型的建立，映射空间维数的确定，自动映射的实现等问题；以层次贝叶斯方法、深度学习方法和增量式方法建立不同任务间的抽象知识表示模型；采用信息熵、最小描述长度准则及近邻传播方法来确定特征维数；拟用多变量IB方法和矩阵论方法处理自动映射问题；项目拟采用不同的迁移学习模型，进行跨媒体模式识别和自动相互注解的研究，力图发现迁移学习所适应的问题特征及数学规律。该项目的研究，一方面有利于设计更加拟人化的机器学习算法，为人工智能的发展提供新的思路和途径，另一方面将迁移学习方法用于海量数据分析，跨媒体模式识别，从而为建立动态快速的决策系统提供坚实的基础。

本项目旨在采用迁移学习思想和方法解决大规模多媒体数据的融合学习问题。采用IB方法实现了多特征和多视角学习，获得了新的聚类算法；采用示范学习方法，实现了快速规划虚拟人物的行为；针对大规模数据，采用矩阵方法获得了相似矩阵的高效计算方法，解决了AP等方法面临的基础性问题之一；采用特征加权设计新型机器学习方法；采用PLSA (Probabilistic Latent Semantic Analysis)方法，调整相关参数，解决聚类结果较为紊乱的问题；采用高维数据划分方法解决此类数据的计算代价问题。上述研究成果解决了多媒体数据迁移学习的一些基础性问题，为进一步实现复杂任务间的知识迁移，大规模数据的增量式学习等提供了可行性。迁移学习具有普遍性，相应的工作也体现在实际应用和教学工作之中。总之，项目组提出了新的机器学习方法及算法，以解决多媒体数据挖掘中的一些关键的或者新颖性的问题，相关研究也体现在了实际应用（河南的地方特色病食管癌等）和教学工作之中，表明了本研究的学术价值和应用价值。