循环神经网络多模态深度模型联想记忆功能研究

项目名称： 循环神经网络多模态深度模型联想记忆功能研究

项目编号： No.61773385

项目类型： 面上项目

立项/批准年度： 2018

项目学科： 其他

项目作者： 杨刚

作者单位： 中国人民大学

项目金额： 16万元

中文摘要： 循环神经网络深度模型在模拟人脑联想记忆功能方面具有重要的科学研究意义。当前深度学习已有效地模拟了人脑区域化分层学习的特点，而循环神经网络深度模型研究方兴未艾，仍有大量问题亟待解决。本课题拟从深度模型模拟人脑联想记忆功能的新角度，研究循环神经网络多模态深度模型及其特性。首先，通过将高效的循环神经网络嵌入到深度学习模型，研究面向联想记忆功能的深度循环神经网络的模型构建和结构优化，并通过引入循环神经网络混沌动态，研究深度循环神经网络的混沌特性和混沌控制，形成多模态下深度循环神经网络的特性对比分析；其次，通过拓展阈限定、贪心分层无监督学习、群智能和演化算法融合等方法，微调网络权值，优化提升深度循环神经网络的联想记忆功能，并展开此模型在视频图像分类问题中的应用研究。本课题从研究人工神经网络模型结构和特点的角度探索人脑联想记忆机理，将为类脑的信息检索提供方法和应用支持。

中文关键词： 循环神经网络；联想记忆；多层次结构与深度学习网络；网络结构学习；深度模型

英文摘要： Recurrent neural network with deep architecture (DCNN) owns important scientific value to explore and mimic brain associative memory mechanism. Now deep learning has already revealed brain property of domain layer-wise learning, but this researching of DCNN on associative memory is just beginning, containing many problems to be solved. From a novel aspect of using deep architecture to mimic brain associative memory, this project studies models, principles and optimizations of DCNN. Firstly, we propose a basic multi-layer associative memory model with different inputting, through embedding cascaded shallow recurrent model and layer-wise training method. We advance the associative memory function by introducing chaotic dynamics and chaos control, further research the model complex dynamics experimentally. Secondly, we extend threshold and regularization method, introduce greedy layer-wise un-surpervised learning, and integrate swarm intelligence to optimize network weight learning. Through implementation of the improved associative memory function, we research DCNN applications on the field of image recognition. Our project could explore the memory mechanism of human brain from the aspect of artificial neural network structure, and it would provide methodology and application supports for constructing brain-like information retrieval.

英文关键词： recurrent neural network;associative memory;multiple layers and deep learning;network structure learning;deep architecture