IDEAS home Printed from https://ideas.repec.org/a/nat/natcom/v16y2025i1d10.1038_s41467-025-58532-9.html
   My bibliography  Save this article

Optimization on multifractal loss landscapes explains a diverse range of geometrical and dynamical properties of deep learning

Author

Listed:
  • Andrew Ly

    (University of Sydney)

  • Pulin Gong

    (University of Sydney)

Abstract

Gradient descent and its variants are foundational in solving optimization problems across many disciplines. In deep learning, these optimizers demonstrate a remarkable ability to dynamically navigate complex loss landscapes, ultimately converging to solutions that generalize well. To elucidate the mechanism underlying this ability, we introduce a theoretical framework that models the complexities of loss landscapes as multifractal. Our model unifies and explains a broad range of realistic geometrical signatures of loss landscapes, including clustered degenerate minima, multiscale structure, and rich optimization dynamics in deep neural networks, such as the edge of stability, non-stationary anomalous diffusion, and the extended edge of chaos without requiring fine-tuning parameters. We further develop a fractional diffusion theory to illustrate how these optimization dynamics, coupled with multifractal structure, effectively guide optimizers toward smooth solution spaces housing flatter minima, thus enhancing generalization. Our findings suggest that the complexities of loss landscapes do not hinder optimization; rather, they facilitate the process. This perspective not only has important implications for understanding deep learning but also extends potential applicability to other disciplines where optimization unfolds on complex landscapes.

Suggested Citation

  • Andrew Ly & Pulin Gong, 2025. "Optimization on multifractal loss landscapes explains a diverse range of geometrical and dynamical properties of deep learning," Nature Communications, Nature, vol. 16(1), pages 1-14, December.
  • Handle: RePEc:nat:natcom:v:16:y:2025:i:1:d:10.1038_s41467-025-58532-9
    DOI: 10.1038/s41467-025-58532-9
    as

    Download full text from publisher

    File URL: https://www.nature.com/articles/s41467-025-58532-9
    File Function: Abstract
    Download Restriction: no

    File URL: https://libkey.io/10.1038/s41467-025-58532-9?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
    ---><---

    References listed on IDEAS

    as
    1. Patrick Charbonneau & Jorge Kurchan & Giorgio Parisi & Pierfrancesco Urbani & Francesco Zamponi, 2014. "Fractal free energy landscapes in structural glasses," Nature Communications, Nature, vol. 5(1), pages 1-6, September.
    Full references (including those not matched with items on IDEAS)

    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. Kyeyune-Nyombi, Eru & Morone, Flaviano & Liu, Wenwei & Li, Shuiqing & Gilchrist, M. Lane & Makse, HernĂ¡n A., 2018. "High-resolution of particle contacts via fluorophore exclusion in deep-imaging of jammed colloidal packings," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 490(C), pages 1387-1395.
    2. Leo Zella & Jaeyun Moon & Takeshi Egami, 2024. "Ripples in the bottom of the potential energy landscape of metallic glass," Nature Communications, Nature, vol. 15(1), pages 1-7, December.
    3. Archana, G.R. & Barik, Debashis, 2024. "Multiple current reversals in driven inertial coupled Brownian particles under rough symmetric periodic potential," Physica A: Statistical Mechanics and its Applications, Elsevier, vol. 650(C).
    4. Yue Deng & Deng Pan & Yuliang Jin, 2024. "Jamming is a first-order transition with quenched disorder in amorphous materials sheared by cyclic quasistatic deformations," Nature Communications, Nature, vol. 15(1), pages 1-12, December.
    5. Zhang, Dongjian & Ma, Qihua & Dong, Hailiang & Liao, He & Liu, Xiangyu & Zha, Yibin & Zhang, Xiaoxiao & Qian, Xiaomin & Liu, Jin & Gan, Xuehui, 2023. "Time-delayed feedback bistable stochastic resonance system and its application in the estimation of the Polyester Filament Yarn tension in the spinning process," Chaos, Solitons & Fractals, Elsevier, vol. 168(C).
    6. Sunny Gupta & Xiaochen Yang & Gerbrand Ceder, 2023. "What dictates soft clay-like lithium superionic conductor formation from rigid salts mixture," Nature Communications, Nature, vol. 14(1), pages 1-8, December.

    More about this item

    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:nat:natcom:v:16:y:2025:i:1:d:10.1038_s41467-025-58532-9. 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: Sonal Shukla or Springer Nature Abstracting and Indexing (email available below). General contact details of provider: http://www.nature.com .

    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.