摘要
笔者通过研究向量来表述一个神经元,由激活向量反映某种实体的特征。笔者展示几何神经网络在小型分类数据集Iris和MNIST上快速训练和高准确率的表现,相比传统BP神经网络模型更易训练,效果更好。距离衰减机制从物理层面描述了神经元之间信息传导的衰弱过程,同时也刻画了神经元与神经元之间的物理结构信息。
This paper studies the use of vectors to describe a neuron, whose activation vector reflects the characteristics of some entity. The paper shows the performance of geometric neural network for fast training and high accuracy in small classified data sets, Iris and MNIST, compared with the traditional BP neural network model, which is easier to train and has better effect. The distance attenuation mechanism describes the weak process of information conduction between neurons from the physical level, and also depicts the physical structure information between neurons and neuron.
作者
龚轩涛
陈昌平
Gong Xuantao;Chen Changping(Tianfu College of Southwestern University of Finance and Economics,Mianyang Sichuan 621000,China)
出处
《信息与电脑》
2018年第19期83-84,87,共3页
Information & Computer
关键词
神经元
神经网络
向量化
距离衰减
neuron
neural network
vectorization
distance attenuation
