IDEAS home Printed from https://ideas.repec.org/a/eee/phsmap/v675y2025ics037843712500487x.html

Low-latency vision transformers via large-scale multi-head attention

Author

Listed:
  • Gross, Ronit D.
  • Halevi, Tal
  • Koresh, Ella
  • Tzach, Yarden
  • Kanter, Ido

Abstract

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.

Suggested Citation

  • Gross, Ronit D. & Halevi, Tal & Koresh, Ella & Tzach, Yarden & Kanter, Ido, 2025. "Low-latency vision transformers via large-scale multi-head attention," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 675(C).
  • Handle: RePEc:eee:phsmap:v:675:y:2025:i:c:s037843712500487x
    DOI: 10.1016/j.physa.2025.130835
    as

    Download full text from publisher

    File URL: http://www.sciencedirect.com/science/article/pii/S037843712500487X
    Download Restriction: Full text for ScienceDirect subscribers only. Journal offers the option of making the article available online on Science direct for a fee of $3,000

    File URL: https://libkey.io/10.1016/j.physa.2025.130835?utm_source=ideas
    LibKey link: if access is restricted and if your library uses this service, LibKey will redirect you to where you can use your library subscription to access this item
    ---><---

    As the access to this document is restricted, you may want to

    for a different version of it.

    References listed on IDEAS

    as
    1. Li, Tianya & Wang, Kejian & Wang, Jihao & Liu, Yueqi & Han, Yufen & Xu, Zhiyang & Lin, Guangyi & Liu, Yong, 2021. "Optimization of GDL to improve water transferability," Renewable Energy, Elsevier, vol. 179(C), pages 2086-2093.
    2. Tevet, Ofek & Gross, Ronit D. & Hodassman, Shiri & Rogachevsky, Tal & Tzach, Yarden & Meir, Yuval & Kanter, Ido, 2024. "Efficient shallow learning mechanism as an alternative to deep learning," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 635(C).
    3. Asim Patra & Mohammed K. A. Kaabar & Sergejs Solovjovs, 2021. "Catalan Transform of k-Balancing Sequences," International Journal of Mathematics and Mathematical Sciences, Hindawi, vol. 2021, pages 1-6, December.
    4. Asim Patra & Mohammed K. A. Kaabar, 2021. "Catalan Transform of k‐Balancing Sequences," International Journal of Mathematics and Mathematical Sciences, John Wiley & Sons, vol. 2021(1).
    5. , Darmadi & Sari, Ratna, 2021. "Gaya Kepemimpinan Transformasional dan Motivasi Kerja," Thesis Commons 8zeh9, Center for Open Science.
    6. Koresh, Ella & Halevi, Tal & Meir, Yuval & Dilmoney, Dolev & Dror, Tamar & Gross, Ronit & Tevet, Ofek & Hodassman, Shiri & Kanter, Ido, 2024. "Scaling in Deep and Shallow Learning Architectures," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 646(C).
    7. Jonah Busch & Irene Ring & Monique Akullo & Oyut Amarjargal & Maud Borie & Rodrigo S. Cassola & Annabelle Cruz-Trinidad & Nils Droste & Joko Tri Haryanto & Ulan Kasymov & Nataliia Viktorivna Kotenko &, 2021. "A global review of ecological fiscal transfers," Nature Sustainability, Nature, vol. 4(9), pages 756-765, September.
    8. Chi Wing Chu & Tony Sit & Gongjun Xu, 2021. "Transformed Dynamic Quantile Regression on Censored Data," Journal of the American Statistical Association, Taylor & Francis Journals, vol. 116(534), pages 874-886, April.
    9. Kun Wang & Christopher W. Johnson & Kane C. Bennett & Paul A. Johnson, 2021. "Predicting fault slip via transfer learning," Nature Communications, Nature, vol. 12(1), pages 1-11, December.
    10. Marcelle Chauvet & Rafael R. S. Guimaraes, 2021. "Transfer Learning for Business Cycle Identification," Working Papers Series 545, Central Bank of Brazil, Research Department.
    11. Zebin Hu & Hao Liu & Zhendong Li & Zekuan Yu & Long Wang, 2021. "Cross-Model Transformer Method for Medical Image Synthesis," Complexity, Hindawi, vol. 2021, pages 1-7, October.
    12. Kung-Jeng Wang & Diwanda Ageng Rizqi & Hong-Phuc Nguyen, 2021. "Skill transfer support model based on deep learning," Journal of Intelligent Manufacturing, Springer, vol. 32(4), pages 1129-1146, April.
    13. Koresh, Ella & Gross, Ronit D. & Meir, Yuval & Tzach, Yarden & Halevi, Tal & Kanter, Ido, 2025. "Unified CNNs and transformers underlying learning mechanism reveals multi-head attention modus vivendi," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 666(C).
    Full references (including those not matched with items on IDEAS)

    Citations

    Citations are extracted by the CitEc Project, subscribe to its RSS feed for this item.
    as


    Cited by:

    1. Gross, Ronit D. & Tzach, Yarden & Halevi, Tal & Koresh, Ella & Kanter, Ido, 2026. "Tiny language models," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 681(C).
    2. Tzach, Yarden & Gross, Ronit D. & Koresh, Ella & Rosner, Shalom & Shpringer, Or & Halevi, Tal & Kanter, Ido, 2026. "Learning mechanism underlying NLP pre-training and fine-tuning," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 682(C).
    3. Rosner, Shalom & Gross, Ronit D. & Koresh, Ella & Kanter, Ido, 2026. "Single-nodal spontaneous symmetry breaking in NLP models," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 688(C).

    Most related items

    These are the items that most often cite the same works as this one and are cited by the same works as this one.
    1. Gou, Liangjie & Yang, Zhaozhong & Min, Chao & Yi, Duo & Li, Xiaogang & Kong, Bing, 2024. "A novel domain adaptation method with physical constraints for shale gas production forecasting," Applied Energy, Elsevier, vol. 371(C).
    2. Li, Dan & Li, Yijun & Wang, Chaoqun & Chen, Min & Wu, Qi, 2023. "Forecasting carbon prices based on real-time decomposition and causal temporal convolutional networks," Applied Energy, Elsevier, vol. 331(C).
    3. Habibi, Mahyar & Hovy, Dirk & Schwarz, Carlo, 2026. "The Content Moderator's Dilemma: Removal of Toxic Content and Distortions to Online Discourse," CEPR Discussion Papers 21257, Centre for Economic Policy Research.
    4. Mahyar Habibi & Dirk Hovy & Carlo Rasmus Schwarz, 2026. "The Content Moderator's Dilemma: Removal of Toxic Content and Distortions to Online Discourse," CESifo Working Paper Series 12521, CESifo.
    5. Yang, Mao & Guo, Zhenpeng & Wang, Da & Wang, Bo & Wang, Zhao & Huang, Tao, 2026. "Short-term photovoltaic power forecasting method considering historical information reuse and numerical weather forecasting," Renewable Energy, Elsevier, vol. 256(PB).
    6. Tzach, Yarden & Gross, Ronit D. & Koresh, Ella & Rosner, Shalom & Shpringer, Or & Halevi, Tal & Kanter, Ido, 2026. "Learning mechanism underlying NLP pre-training and fine-tuning," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 682(C).
    7. Hocke, Simone & Klee, Andreas, 2023. "Transformation in der Arbeitswelt gestalten: Welchen Beitrag leistet eine akademische Weiterbildung von Betriebs- und Personalräten?," Working Paper Forschungsförderung 309, Hans-Böckler-Stiftung, Düsseldorf.
    8. Kateryna Onopriienko & Tetyana Vasylieva & Salmane Bourekkadi & Robert Rehak & Nadiia Artyukhova, 2023. "Effectiveness of open education in the lifelong learning system," Philosophy, Economics and Law Review Articles, Philosophy, Economics and Law Review, vol. 3(2), pages 103-111, December.
    9. Meir, Yuval & Tevet, Ofek & Tzach, Yarden & Hodassman, Shiri & Kanter, Ido, 2024. "Role of delay in brain dynamics," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 654(C).
    10. Gross, Ronit D. & Tzach, Yarden & Halevi, Tal & Koresh, Ella & Kanter, Ido, 2026. "Tiny language models," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 681(C).
    11. Gross, Ronit & Koresh, Ella & Halevi, Tal & Hodassman, Shiri & Meir, Yuval & Tzach, Yarden & Kanter, Ido, 2025. "Multilabel classification outperforms detection-based technique," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 658(C).
    12. Christopher W. Johnson & Kun Wang & Paul A. Johnson, 2025. "Automatic speech recognition predicts contemporaneous earthquake fault displacement," Nature Communications, Nature, vol. 16(1), pages 1-10, December.
    13. Qi, Yu & Yin, Aoxue & Chen, Jianwei & Yang, Chunfei & Zhan, Pengyu, 2024. "Motivating for environmental protection: Evidence from county officials in China," World Development, Elsevier, vol. 184(C).
    14. Liu, Huize & Hu, Zunyan & Li, Jianqiu & Xu, Liangfei & Shao, Yangbin & Ouyang, Minggao, 2023. "Investigation on the optimal GDL thickness design for PEMFCs considering channel/rib geometry matching and operating conditions," Energy, Elsevier, vol. 282(C).
    15. Huang, Dayan & Kou, Aiju & Liu, Chengyi & Liu, Shanmin, 2023. "The effect of PWS arrangements on M&A activities," Finance Research Letters, Elsevier, vol. 52(C).
    16. Sean Dougherty & Andoni Montes Nebreda, 2022. "Going global, locally? Decentralized environmental expenditure and air quality," Public Sector Economics, Institute of Public Finance, vol. 46(4), pages 489-503.
    17. Ren, Shenggang & Zhou, Qiong & Zhang, Xinxin & Zeng, Huixiang, 2024. "How do heavily polluting firms cope with dual environmental regulation? A study from the perspective of financial asset allocation," Energy Economics, Elsevier, vol. 139(C).
    18. Ajidarma, Praditya & Nof, Shimon Y., 2025. "Skill-and-Knowledge Sharing HUB-CI model for resilient production systems," International Journal of Production Economics, Elsevier, vol. 287(C).
    19. Tan Ching Ng & Sie Yee Lau & Morteza Ghobakhloo & Masood Fathi & Meng Suan Liang, 2022. "The Application of Industry 4.0 Technological Constituents for Sustainable Manufacturing: A Content-Centric Review," Sustainability, MDPI, vol. 14(7), pages 1-21, April.
    20. Qiao, Zhenhua & Xu, Xinyi & Zou, Weitao & Huang, Yingli, 2024. "Urban sustainable development goals and ecosystem services: Pathways to achieving coordination," Land Use Policy, Elsevier, vol. 146(C).

    More about this item

    Keywords

    ;
    ;
    ;
    ;

    Statistics

    Access and download statistics

    Corrections

    All material on this site has been provided by the respective publishers and authors. You can help correct errors and omissions. When requesting a correction, please mention this item's handle: RePEc:eee:phsmap:v:675:y:2025:i:c:s037843712500487x. See general information about how to correct material in RePEc.

    If you have authored this item and are not yet registered with RePEc, we encourage you to do it here. This allows to link your profile to this item. It also allows you to accept potential citations to this item that we are uncertain about.

    If CitEc recognized a bibliographic reference but did not link an item in RePEc to it, you can help with this form .

    If you know of missing items citing this one, you can help us creating those links by adding the relevant references in the same way as above, for each refering item. If you are a registered author of this item, you may also want to check the "citations" tab in your RePEc Author Service profile, as there may be some citations waiting for confirmation.

    For technical questions regarding this item, or to correct its authors, title, abstract, bibliographic or download information, contact: Catherine Liu (email available below). General contact details of provider: http://www.journals.elsevier.com/physica-a-statistical-mechpplications/ .

    Please note that corrections may take a couple of weeks to filter through the various RePEc services.

    IDEAS is a RePEc service. RePEc uses bibliographic data supplied by the respective publishers.