A. V. Trusov, E. E. Limonova, D. P. Nikolaev, V. V. Arlazarov
{"title":"On Fast Computing of Neural Networks Using Central Processing Units","authors":"A. V. Trusov, E. E. Limonova, D. P. Nikolaev, V. V. Arlazarov","doi":"10.1134/s105466182304048x","DOIUrl":null,"url":null,"abstract":"<h3 data-test=\"abstract-sub-heading\">Abstract</h3><p>This work is devoted to methods for creating fast and accurate neural network algorithms for central processors, which were proposed by scientists of the V.L. Arlazarov’s scientific school. It outlines general principles and approaches to improving computational efficiency and discusses specific examples: tensor convolution decompositions that simplify convolutional neural networks; bounded nonlinear activation function ratio, which is calculated faster than exponential activation functions; and p-im2col convolution algorithm, which allows you to achieve a balance between computational efficiency and RAM consumption. Particular attention is paid to quantized (8- and 4-bit integer) neural networks, their training, implementation, and limitations on some central processor architectures, such as Elbrus.</p>","PeriodicalId":35400,"journal":{"name":"PATTERN RECOGNITION AND IMAGE ANALYSIS","volume":"28 1","pages":""},"PeriodicalIF":0.7000,"publicationDate":"2024-03-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"PATTERN RECOGNITION AND IMAGE ANALYSIS","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1134/s105466182304048x","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q4","JCRName":"COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS","Score":null,"Total":0}
引用次数: 0
Abstract
This work is devoted to methods for creating fast and accurate neural network algorithms for central processors, which were proposed by scientists of the V.L. Arlazarov’s scientific school. It outlines general principles and approaches to improving computational efficiency and discusses specific examples: tensor convolution decompositions that simplify convolutional neural networks; bounded nonlinear activation function ratio, which is calculated faster than exponential activation functions; and p-im2col convolution algorithm, which allows you to achieve a balance between computational efficiency and RAM consumption. Particular attention is paid to quantized (8- and 4-bit integer) neural networks, their training, implementation, and limitations on some central processor architectures, such as Elbrus.
期刊介绍：
The purpose of the journal is to publish high-quality peer-reviewed scientific and technical materials that present the results of fundamental and applied scientific research in the field of image processing, recognition, analysis and understanding, pattern recognition, artificial intelligence, and related fields of theoretical and applied computer science and applied mathematics. The policy of the journal provides for the rapid publication of original scientific articles, analytical reviews, articles of the world''s leading scientists and specialists on the subject of the journal solicited by the editorial board, special thematic issues, proceedings of the world''s leading scientific conferences and seminars, as well as short reports containing new results of fundamental and applied research in the field of mathematical theory and methodology of image analysis, mathematical theory and methodology of image recognition, and mathematical foundations and methodology of artificial intelligence. The journal also publishes articles on the use of the apparatus and methods of the mathematical theory of image analysis and the mathematical theory of image recognition for the development of new information technologies and their supporting software and algorithmic complexes and systems for solving complex and particularly important applied problems. The main scientific areas are the mathematical theory of image analysis and the mathematical theory of pattern recognition. The journal also embraces the problems of analyzing and evaluating poorly formalized, poorly structured, incomplete, contradictory and noisy information, including artificial intelligence, bioinformatics, medical informatics, data mining, big data analysis, machine vision, data representation and modeling, data and knowledge extraction from images, machine learning, forecasting, machine graphics, databases, knowledge bases, medical and technical diagnostics, neural networks, specialized software, specialized computational architectures for information analysis and evaluation, linguistic, psychological, psychophysical, and physiological aspects of image analysis and pattern recognition, applied problems, and related problems. Articles can be submitted either in English or Russian. The English language is preferable. Pattern Recognition and Image Analysis is a hybrid journal that publishes mostly subscription articles that are free of charge for the authors, but also accepts Open Access articles with article processing charges. The journal is one of the top 10 global periodicals on image analysis and pattern recognition and is the only publication on this topic in the Russian Federation, Central and Eastern Europe.
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号
