统计认知分类/聚类模型及其模型计算方法研究

项目名称： 统计认知分类/聚类模型及其模型计算方法研究

项目编号： No.60872071

项目类型： 面上项目

立项/批准年度： 2009

项目学科： 生物科学

项目作者： 钱沄涛

作者单位： 浙江大学

项目金额： 30万元

中文摘要： 统计认知理论把概率统计方法和认知过程相结合，为发展先进智能信息处理技术提供了新的途径。本项目在机器学习、模式识别和多媒体信息处理的研究成果基础上，针对信息分类/聚类这一人类认知和智能信息处理的基础性和关键性问题，重点研究统计认知分类/聚类模型和模型计算方法。项目主要工作包括：1）讨论了基于概率图模型和稀疏模型的数据结构特征表达和处理理论，可以有效模拟随机性、不确定性和稀疏性等认知行为，构造了融合多结构特征的统计认知模型框架，提出了基于稀疏盲信号分解的数据降维算法、基于图结构和统计推理的聚类算法，基于固定基和自适应字典的稀疏线性和非线性回归和分类算法，结构高斯过程回归和分类算法等。2）研究了近似推理、模块分解、凸优化等统计认知模型计算方法，提出了基于稀疏学习的模型参数和结构简化算法、基于信度传播的统计推理算法、和基于凸优化的模型参数估计方法等。3）把统计认知模型应用于在线生物医学文献信息处理和高光谱成像信息处理，取得了非常好的效果。总体上本项目研究处于国际前沿水平。

中文关键词： 统计认知理论；分类/聚类；概率图模型；稀疏建模

英文摘要： Statistical cognition theory combines statistical methods and cognitive process models, which provides a new way for developing advanced intelligent information processing methods. Focused on classification/clustering that is one of the foundmental problems of human cognition and intelligent information processing, statistical cognitive classification and clustering models and their computation methods are studied. The main work includes: 1) exploiting structural feature representation and processing methods based on probabilistic graphical model and sparse modeling, which can implement some cognitive behaviors such as randomness, uncertainty and sparsity; building multi-structural statistical cognition framework; proposing graph structure and statistical inference based clustering algorithm, fixed or adaptive dictionary based linear/nonlinear sparse regression and classification algorithms, and structured gaussian process regression and classification algorithms. 2) studying some important computational methods including approximate inference, modular decomposition, and convex optimization; proposing sparse learning based parameter and structure simplification method, belief-propagation based statistical inference method, and convex optimization based parameter estimation method. 3) applying statistical cognitive models for online biological journal article analysis and hyperspectral imaging data processing, and demonstrating the effectiveness and power of our methods. As a whole, the technical level of this project can be compared with that of other research groups in the world.

英文关键词： Statistical cognition theory; Classification/clustering; Probabilistic graph models; sparse modeling