不确定性大数据的高效学习模型与算法，是当前大数据智能处理研究的热点关键问题。针对传统大数据机器学习“由细到粗”的信息处理模式与人类智能“由粗到细”的认知处理模式之间的矛盾，本项目借鉴人类智能问题求解的多粒度思维机制，采用人类大脑“大范围优先”的认知机制和智能控制系统中“计算前置”的信息处理机制，融合知识驱动与数据驱动的计算模式，研究大数据的多粒度学习模型理论与方法，具体包括：从粗粒度到细粒度逐步细化的多粒度渐进式分解求解、从细粒度到粗粒度逐步扩张的多粒度局部向全局的扩张优化计算、融合知识驱动和数据驱动的双向多粒度智能计算、基于粒计算的多粒度深度神经网络学习等4个方面的研究内容。本项目研究将建立满足时限约束条件的多维度跨领域大数据的多粒度计算模型，为解决深度神经网络的结构设计问题和解释困难问题，提供理论基础和技术方法。

It is one of the key issues of big data intelligent processing to develop efficient learning models and algorithms for uncertain big data. There is a contradiction between the "from fine to coarse" machine information processing model of traditional big data machine learning and the "from coarse to fine" human cognitive processing model. In view of the multi-granularity thinking mechanism of human intelligence problem solving, using the human cognition mechanism of "global precedence" and the information processing mechanism of "prefix computation" of intelligent control systems, this project studies the multi-granularity learning model and methods for big data based on integrating two computation models, knowledge driven and data driven, to solve this contradiction problem. There are 4 major research tasks, that is, the multi-granularity progressive decomposition problem solving mechanism of “from coarse to fine”, the multi-granularity expansion optimization computing mechanism of “from fine to coarse”, the bi-directional multi-granularity intelligent computing mechanism based on the integration of knowledge driven and data driven, and the multi-granularity deep neural networks based on granular computing. A multi-granularity computing model with time constraints for multi-dimension and cross-domain big data will be developed. It could provide theoretical basis and technical methods for solving the problems of structural design and interpretation difficult of deep neural networks.