【作者】 杨静； 谢昭； 孙永宣；

【Author】 Yang Jing Xie Zhao Sun Yongxuan (Department of Computer and Information Science Hefei University of Technology Hefei 230009 China)

【机构】 合肥工业大学计算机与信息学院图像信息处理研究室；

【摘要】 本文借鉴人类记忆和模式识别智能行为易于融入模式几何结构信息和生物神经系统本身连接非均匀性和稀疏性的内在特点,考虑从神经元互联结构入手,将神经元间的结构化互联信息测度引入传统全互联的Hopfield网络中,构建了一种结构化联想记忆模型(Contextual Hopfield Neural Networks,CHNN)。同时以概率统计分析为手段,开展了对基于Hebb规则的CHNN模型联想记忆性能的理论研究。理论分析表明:无论何种形式的结构化互联信息的引入对于以相关学习Hebb规则作为权重学习算法的网络模型而言,无益于网络联想记忆性能的增强,大量数值仿真实验也验证了该理论分析结论。将此结论引申出去,为后续探讨网络实现稀疏互联的可能性提供了重要启示。更多还原

【Abstract】 A novel Hopfield neural networks based on contextual information is proposed in this paper.Aimed at imitating the internal characteristic of human intelligence of memory and recognition through embedding more geometrical contextual information,and takes the heterogeneous together with sparse of connections in nature biological neural system into account,this new model introduce contextual information into the connection of neurons.Using probabilistic approach,associative memory ability of the new model based on Hebb rule is analyzed in detail.We conclude that the introduction of the contextual information has no beneficial effect;still the implications for exploring the sparse topology of the networks are suggestive.更多还原