Tackling over-smoothing in multi-label image classification using graphical convolution neural network.

IF 2.7 4区 计算机科学 Q3 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE
Vikas Chauhan, Aruna Tiwari, Boppudi Venkata, Vislavath Naik
{"title":"Tackling over-smoothing in multi-label image classification using graphical convolution neural network.","authors":"Vikas Chauhan, Aruna Tiwari, Boppudi Venkata, Vislavath Naik","doi":"10.1007/s12530-022-09463-z","DOIUrl":null,"url":null,"abstract":"<p><p>The importance of the graphical convolution network in multi-label classification has grown in recent years due to its label embedding representation capabilities. The graphical convolution network is able to capture the label dependencies using the correlation between labels. However, the graphical convolution network suffers from an over-smoothing problem when the layers are increased in the network. Over-smoothing makes the nodes indistinguishable in the deep graphical convolution network. This paper proposes a normalization technique to tackle the over-smoothing problem in the graphical convolution network for multi-label classification. The proposed approach is an efficient multi-label object classifier based on a graphical convolution neural network that tackles the over-smoothing problem. The proposed approach normalizes the output of the graph such that the total pairwise squared distance between nodes remains the same after performing the convolution operation. The proposed approach outperforms the existing state-of-the-art approaches based on the results obtained from the experiments performed on MS-COCO and VOC2007 datasets. The experimentation results show that pairnorm mitigates the effect of over-smoothing in the case of using a deep graphical convolution network.</p>","PeriodicalId":12174,"journal":{"name":"Evolving Systems","volume":"44 1","pages":"1-11"},"PeriodicalIF":2.7000,"publicationDate":"2022-09-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9451128/pdf/","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Evolving Systems","FirstCategoryId":"94","ListUrlMain":"https://doi.org/10.1007/s12530-022-09463-z","RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q3","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0

Abstract

The importance of the graphical convolution network in multi-label classification has grown in recent years due to its label embedding representation capabilities. The graphical convolution network is able to capture the label dependencies using the correlation between labels. However, the graphical convolution network suffers from an over-smoothing problem when the layers are increased in the network. Over-smoothing makes the nodes indistinguishable in the deep graphical convolution network. This paper proposes a normalization technique to tackle the over-smoothing problem in the graphical convolution network for multi-label classification. The proposed approach is an efficient multi-label object classifier based on a graphical convolution neural network that tackles the over-smoothing problem. The proposed approach normalizes the output of the graph such that the total pairwise squared distance between nodes remains the same after performing the convolution operation. The proposed approach outperforms the existing state-of-the-art approaches based on the results obtained from the experiments performed on MS-COCO and VOC2007 datasets. The experimentation results show that pairnorm mitigates the effect of over-smoothing in the case of using a deep graphical convolution network.

利用图形卷积神经网络解决多标签图像分类中的过度平滑问题。
近年来,由于其标签嵌入表示能力,图形卷积网络在多标签分类中的重要性越来越大。图形卷积网络能够利用标签之间的相关性来捕获标签依赖关系。然而,随着网络层数的增加,图形卷积网络存在过度平滑问题。在深度图形卷积网络中,过度平滑使得节点难以区分。本文提出了一种归一化技术来解决图形卷积网络中多标签分类的过度平滑问题。该方法是一种基于图形卷积神经网络的高效多标签目标分类器,解决了过度平滑问题。所提出的方法将图的输出归一化,使得节点之间的总成对平方距离在执行卷积操作后保持不变。基于MS-COCO和VOC2007数据集的实验结果,本文提出的方法优于现有的最先进的方法。实验结果表明,在使用深度图形卷积网络的情况下,对范数减轻了过度平滑的影响。
本文章由计算机程序翻译,如有差异,请以英文原文为准。
求助全文
约1分钟内获得全文 求助全文
来源期刊
Evolving Systems
Evolving Systems COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE-
CiteScore
7.80
自引率
6.20%
发文量
67
期刊介绍: Evolving Systems covers surveys, methodological, and application-oriented papers in the area of dynamically evolving systems. ‘Evolving systems’ are inspired by the idea of system model evolution in a dynamically changing and evolving environment. In contrast to the standard approach in machine learning, mathematical modelling and related disciplines where the model structure is assumed and fixed a priori and the problem is focused on parametric optimisation, evolving systems allow the model structure to gradually change/evolve. The aim of such continuous or life-long learning and domain adaptation is self-organization. It can adapt to new data patterns, is more suitable for streaming data, transfer learning and can recognise and learn from unknown and unpredictable data patterns. Such properties are critically important for autonomous, robotic systems that continue to learn and adapt after they are being designed (at run time). Evolving Systems solicits publications that address the problems of all aspects of system modelling, clustering, classification, prediction and control in non-stationary, unpredictable environments and describe new methods and approaches for their design. The journal is devoted to the topic of self-developing, self-organised, and evolving systems in its entirety — from systematic methods to case studies and real industrial applications. It covers all aspects of the methodology such as Evolving Systems methodology Evolving Neural Networks and Neuro-fuzzy Systems Evolving Classifiers and Clustering Evolving Controllers and Predictive models Evolving Explainable AI systems Evolving Systems applications but also looking at new paradigms and applications, including medicine, robotics, business, industrial automation, control systems, transportation, communications, environmental monitoring, biomedical systems, security, and electronic services, finance and economics. The common features for all submitted methods and systems are the evolving nature of the systems and the environments. The journal is encompassing contributions related to: 1) Methods of machine learning, AI, computational intelligence and mathematical modelling 2) Inspiration from Nature and Biology, including Neuroscience, Bioinformatics and Molecular biology, Quantum physics 3) Applications in engineering, business, social sciences.
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
确定
请完成安全验证×
copy
已复制链接
快去分享给好友吧!
我知道了
右上角分享
点击右上角分享
0
联系我们:info@booksci.cn Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。 Copyright © 2023 布克学术 All rights reserved.
京ICP备2023020795号-1
ghs 京公网安备 11010802042870号
Book学术文献互助
Book学术文献互助群
群 号:604180095
Book学术官方微信