摘要
神经网络虽然在众多模式识别任务中取得了巨大的成功,但是由于其多层的非线性变换结构,使得人们难以直观理解和高效准确地使用这类模型,该问题在当前广泛使用的深度神经网络中表现得尤为突出。可视化技术以其简洁直观的特点成为我们理解复杂模型工作机制的重要手段,这使得神经网络可视化技术成为深度学习中一个学术研究热点。本文着眼于全连接神经网络对原始特征空间的剖分过程,从特征变换、剖分、编码三个角度分析了全连接神经网络在分类过程中的作用。对网络最小的分类单元——胞腔的形成和分解过程进行了分析和可视化。借助本文提出的激活编码的方式,使得我们对于无法直观可视化的高维空间的剖分情况能够进行一定程度的了解,并能够成为定义与讨论“压缩”与“自正则”两种现象的一种工具。通过分析相同训练数据下不同网络结构的表现以及相同网络结构在不同训练数据上的表现,揭示了自正则、胞腔数量以及网络学习能力之间的联系。
Although neural network has achieved great success in many pattern recognition tasks,because of its multi-layer non-linear transformation structure,it is difficult for people to intuitively understand and use this kind of model efficiently and accurately.This problem is particularly prominent in the current widely used deep neural network.Visualization technology,with its concise and intuitive characteristics,has become an important means for us to understand the working mechanism of complex models,which makes the visualization technology of neural networks become a hot academic research topic in deep learning.This paper focuses on the partition process of full-connected neural network to the original feature space,and analyses the role of full-connected neural network in the classification process from three angles of feature transformation,partition and coding.The formation and partitioning process of the cell,the smallest classification unit in the network,is analyzed and visualized.With the method of activation coding proposed in this paper,we can understand the partition of high-dimensional space which can not be visualized intuitively to a certain extent,and become a tool for defining and discussing the two phenomena of“compression”and“self-regularity”.By analyzing the performance of different network structures under the same training data and the performance of the same network structure on different training data,the relationship among self-regularity,cell number and network learning ability is revealed.
作者
纪超
唐宏
Ji Chao;Tang Hong(State Key Laboratory of Remote Sensing Science, Beijing Normal University, Beijing 100875, China)
出处
《信号处理》
CSCD
北大核心
2020年第4期486-494,共9页
Journal of Signal Processing
基金
国家自然科学基金:卷积神经网络的遥感图像时空元数据嵌入方法及应用研究(41971280)。
关键词
神经网络
可视化
特征变换
剖分
编码
neural network
visualization
feature transformation
partition
coding
