Naoki Sato (佐藤 尚樹)

I'm a PhD student at Meiji University in Japan. Feel free to reach out with any comments or questions about our paper and research.

Email  /  CV  /  Google Scholar  /  GitHub

profile photo

Research

I'm interested in Optimization for Machine Learning. In particular, I'm interested in the global optimization of non-convex function in deep learning. Recently, I have also been working on Fixed Point Theory for machine learning.

News

[2026-06-02] Our paper on convergence analysis of SGD with momentum under heavy-tailed noise will appear in UAI2026.
[2026-01-22] Our paper on accelerating deep equilibrium models will appear in AISTATS2026.
[2025-02-25] Our paper on graduated optimization will appear in AAAI-25 and I received $1,500 for travel expenses from the Google Conference Scholarship.

Preprints

Convergence Bound and Critical Batch Size of Muon Optimizer
Naoki Sato, Hiroki Naganuma, Hideaki Iiduka
Preprint

Using Stochastic Gradient Descent to Smooth Nonconvex Functions: Analysis of Implicit Graduated Optimization
Naoki Sato, Hideaki Iiduka
Preprint

Publications Talks Biography

  1. Vanilla SGD with Momentum Survives Heavy-Tailed Noise: Convergence Analysis without Gradient Clipping or Normalization
    Ryusei Yamada, Naoki Sato, Hideaki Iiduka
    Conference on Uncertainty in Artificial Intelligence (UAI), 2026
    arXiv  /  code  /  poster
  2. Lipschitz Multiscale Deep Equilibrium Models: A Theoretically Guaranteed and Accelerated Approach
    Naoki Sato, Hideaki Iiduka
    International Conference on Artificial Intelligence and Statistics (AISTATS), 2026
    arXiv  /  code  /  poster
  3. Explicit and Implicit Graduated Optimization in Deep Neural Networks
    Naoki Sato, Hideaki Iiduka
    AAAI Conference on Artificial Intelligence (AAAI), 2025
    arXiv  /  code  /  poster
  4. Scaled Conjugate Gradient Method for Nonconvex Optimization in Deep Neural Networks
    Naoki Sato, Koshiro Izumi, Hideaki Iiduka
    Journal of Machine Learning Research (JMLR), 2024
    arXiv  /  code
  5. Existence and Estimation of Critical Batch Size for Training Generative Adversarial Networks with Two Time-Scale Update Rule
    Naoki Sato, Hideaki Iiduka
    International Conference on Machine Learning (ICML), 2023
    arXiv  /  code  /  poster

The website source code was stolen from this.