Ronit D. Gross , Tal Halevi , Ella Koresh , Yarden Tzach , Ido Kanter
{"title":"通过大规模多头注意实现的低延迟视觉变形","authors":"Ronit D. Gross , Tal Halevi , Ella Koresh , Yarden Tzach , Ido Kanter","doi":"10.1016/j.physa.2025.130835","DOIUrl":null,"url":null,"abstract":"<div><div>The emergence of spontaneous symmetry breaking among a few heads of multi-head attention (MHA) across transformer blocks in classification tasks was recently demonstrated through the quantification of single-nodal performance (SNP). This finding indicates that each head focuses its attention on a subset of labels through cooperation among its SNPs. This underlying learning mechanism is generalized to large-scale MHA (LS-MHA) using a single matrix value representing single-head performance (SHP), analogous to single-filter performance in convolutional neural networks (CNNs). The results indicate that each SHP matrix comprises multiple unit clusters such that each label being explicitly recognized by a few heads with negligible noise. This leads to an increased signal-to-noise ratio (SNR) along the transformer blocks, thereby improving classification accuracy. These features give rise to several distinct vision transformer (ViT) architectures that achieve the same accuracy but differ in their LS-MHA structures. As a result, their soft committee yields superior accuracy, an outcome not typically observed in CNNs which rely on hundreds of filters. In addition, a significant reduction in latency is achieved without affecting the accuracy by replacing the initial transformer blocks with convolutional layers. This substitution accelerates early-stage learning, which is then improved by subsequent transformer layers. The extension of this learning mechanism to natural language processing tasks, based on quantitative differences between CNNs and ViT architectures, has the potential to yield new insights in deep learning. The findings are demonstrated using compact convolutional transformer architectures trained on the CIFAR-100 dataset.</div></div>","PeriodicalId":20152,"journal":{"name":"Physica A: Statistical Mechanics and its Applications","volume":"675 ","pages":"Article 130835"},"PeriodicalIF":3.1000,"publicationDate":"2025-07-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Low-latency vision transformers via large-scale multi-head attention\",\"authors\":\"Ronit D. Gross , Tal Halevi , Ella Koresh , Yarden Tzach , Ido Kanter\",\"doi\":\"10.1016/j.physa.2025.130835\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>The emergence of spontaneous symmetry breaking among a few heads of multi-head attention (MHA) across transformer blocks in classification tasks was recently demonstrated through the quantification of single-nodal performance (SNP). This finding indicates that each head focuses its attention on a subset of labels through cooperation among its SNPs. This underlying learning mechanism is generalized to large-scale MHA (LS-MHA) using a single matrix value representing single-head performance (SHP), analogous to single-filter performance in convolutional neural networks (CNNs). The results indicate that each SHP matrix comprises multiple unit clusters such that each label being explicitly recognized by a few heads with negligible noise. This leads to an increased signal-to-noise ratio (SNR) along the transformer blocks, thereby improving classification accuracy. These features give rise to several distinct vision transformer (ViT) architectures that achieve the same accuracy but differ in their LS-MHA structures. As a result, their soft committee yields superior accuracy, an outcome not typically observed in CNNs which rely on hundreds of filters. In addition, a significant reduction in latency is achieved without affecting the accuracy by replacing the initial transformer blocks with convolutional layers. This substitution accelerates early-stage learning, which is then improved by subsequent transformer layers. The extension of this learning mechanism to natural language processing tasks, based on quantitative differences between CNNs and ViT architectures, has the potential to yield new insights in deep learning. The findings are demonstrated using compact convolutional transformer architectures trained on the CIFAR-100 dataset.</div></div>\",\"PeriodicalId\":20152,\"journal\":{\"name\":\"Physica A: Statistical Mechanics and its Applications\",\"volume\":\"675 \",\"pages\":\"Article 130835\"},\"PeriodicalIF\":3.1000,\"publicationDate\":\"2025-07-22\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Physica A: Statistical Mechanics and its Applications\",\"FirstCategoryId\":\"101\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S037843712500487X\",\"RegionNum\":3,\"RegionCategory\":\"物理与天体物理\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"PHYSICS, MULTIDISCIPLINARY\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Physica A: Statistical Mechanics and its Applications","FirstCategoryId":"101","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S037843712500487X","RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"PHYSICS, MULTIDISCIPLINARY","Score":null,"Total":0}
Low-latency vision transformers via large-scale multi-head attention
The emergence of spontaneous symmetry breaking among a few heads of multi-head attention (MHA) across transformer blocks in classification tasks was recently demonstrated through the quantification of single-nodal performance (SNP). This finding indicates that each head focuses its attention on a subset of labels through cooperation among its SNPs. This underlying learning mechanism is generalized to large-scale MHA (LS-MHA) using a single matrix value representing single-head performance (SHP), analogous to single-filter performance in convolutional neural networks (CNNs). The results indicate that each SHP matrix comprises multiple unit clusters such that each label being explicitly recognized by a few heads with negligible noise. This leads to an increased signal-to-noise ratio (SNR) along the transformer blocks, thereby improving classification accuracy. These features give rise to several distinct vision transformer (ViT) architectures that achieve the same accuracy but differ in their LS-MHA structures. As a result, their soft committee yields superior accuracy, an outcome not typically observed in CNNs which rely on hundreds of filters. In addition, a significant reduction in latency is achieved without affecting the accuracy by replacing the initial transformer blocks with convolutional layers. This substitution accelerates early-stage learning, which is then improved by subsequent transformer layers. The extension of this learning mechanism to natural language processing tasks, based on quantitative differences between CNNs and ViT architectures, has the potential to yield new insights in deep learning. The findings are demonstrated using compact convolutional transformer architectures trained on the CIFAR-100 dataset.
期刊介绍:
Physica A: Statistical Mechanics and its Applications
Recognized by the European Physical Society
Physica A publishes research in the field of statistical mechanics and its applications.
Statistical mechanics sets out to explain the behaviour of macroscopic systems by studying the statistical properties of their microscopic constituents.
Applications of the techniques of statistical mechanics are widespread, and include: applications to physical systems such as solids, liquids and gases; applications to chemical and biological systems (colloids, interfaces, complex fluids, polymers and biopolymers, cell physics); and other interdisciplinary applications to for instance biological, economical and sociological systems.