神经信息编码中的鲁棒性特征子集选择研究

项目名称： 神经信息编码中的鲁棒性特征子集选择研究

项目编号： No.U1304602

项目类型： 联合基金项目

立项/批准年度： 2014

项目学科： 无线电电子学、电信技术

项目作者： 尚志刚

作者单位： 郑州大学

项目金额： 30万元

中文摘要： 从神经电生理信号中提取和选择具有分类鲁棒性的特征，对于深入理解神经信息编码机制和准确解码具有重要价值。本研究目的：针对从原始检测神经电信号中提取的大量特征，基于样本数据集的局部统计结构分析，从高维特征空间中选择出具有良好分类和强泛化能力的鲁棒性特征子集。研究内容包括：（1）分析基于距离、信息熵、相关性、一致性等不同测度特征选择结果的分类鲁棒性和准确性差异，明确不同测度在特征选择方面的联系与差异；（2）研究样本分布局部统计结构分析方法，挖掘高维特征数据集是否内蕴流形结构信息；（3）基于流形学习获得的特征关系，建立具有最大分类间隔期望的评价函数，构建并完善控制参数自适应、嵌入样本数据集局部统计结构信息的优化算法选择鲁棒性特征子集，利用国际通用特征选择数据集进行特征选择效果评价；（4）开展利用大鼠视皮层微电极阵列检测神经信号分类识别不同视觉刺激的动物实验，完成研究方案的实验评估。

中文关键词： 鲁棒性特征子集；特征选择；优化算法；局部统计结构；神经集群编码

英文摘要： It is important and valuable to extract and select features with classification robustness from neural electrophysiology recoding signal for deeply understanding the mechanism of neural information encoding and correctly decoding it. This research aims to select the robust feature subset which has a good classify capability and better generalization capability from the high-dimensional feature space,based on analysising of local statistic structure of observation dataset, in which each feature is extracted from the raw neural electrical signal.The research content includes as follow:Firstly, in order to know the interaction and difference between different inter-feature metric including distance,information entropy,correlation and consistency ,the the feature selection results using them individually are compared to evaluate both classification robustness and correctness.Secondly,the method to analysis the local statistical structure information of observations distribution is studied to learn whether there are underlying low dimension manifold lie in the high dimension dataset. Furthermore, base on the manifold learning,the fitness funciton to optimize to get the maximum classification margin expection is created, so the optimization algorithm is stuied and improved, in which the control parameter is adaptive

英文关键词： robust feature subset；feature selection；optimization algorithm；local statistical structure；neural population encoding