Publications

A compact list of publications with links to the original research records and related resources.

Google Scholar · ORCID

Journal articles

TMLR 2026
Generalization Measures under Controlled Covariate Shift: A Regime-Aware Benchmark
Sora Nakai, Youssef Fadhloun (equal contribution), Kacem Mathlouthi (equal contribution), Kotaro Yoshida (equal contribution), Ganesh Talluri (equal contribution), Ioannis Mitliagkas, Hiroki Naganuma
Transactions on Machine Learning Research, August 2026 (accepted)
arxiv v1 (previous title)
TMLR 2026
What do near-optimal learning rate schedules look like?
Hiroki Naganuma, Atish Agarwala, Priya Kasimbeg, George E. Dahl
Transactions on Machine Learning Research, June 2026
TMLR 2025
Pseudo-Asynchronous Local SGD: Robust and Efficient Data-Parallel Training
Hiroki Naganuma (equal contribution), Xinzhi Zhang (equal contribution), Man-Chung Yue, Ioannis Mitliagkas, Russell J. Hewett, Philipp Andre Witte, Yin Tat Lee
Transactions on Machine Learning Research, August 2025
arxiv code
TMLR 2025
An Empirical Study of Pre-trained Model Selection for Out-of-Distribution Generalization and Calibration
Hiroki Naganuma (equal contribution), Ryuichiro Hataya (equal contribution), Kotaro Yoshida, Ioannis Mitliagkas
Transactions on Machine Learning Research, April 2025
code
PRL
Geometric Insights into Focal Loss: Reducing Curvature for Enhanced Model Calibration
Masanari Kimura (equal contribution), Hiroki Naganuma (equal contribution)
Pattern Recognition Letters, February 2025
TMLR 2024
Towards Understanding Variants of Invariant Risk Minimization from the Perspective of Calibration
Kotaro Yoshida (equal contribution), Hiroki Naganuma (equal contribution)
Transactions on Machine Learning Research, June 2024
code
TMLR 2023
Empirical Study on Optimizer Selection for Out-of-Distribution Generalizations
Hiroki Naganuma, Kartik Ahuja, Shiro Takagi, Tetsuya Motokawa, Rio Yokota, Kohta Ishikawa, Ikuro Sato, Ioannis Mitliagkas
Transactions on Machine Learning Research, June 2023
code

Conference papers

ICML 2026
Adaptive Batch Sizes Using Non-Euclidean Gradient Noise Scales for Stochastic Sign and Spectral Descent
Hiroki Naganuma, Shagun Gupta, Youssef Briki, Ioannis Mitliagkas, Irina Rish, Parameswaran Raman, Hao-Jun Michael Shi
International Conference on Machine Learning, July 2026
ICLR 2026
On Fairness of Task Arithmetic: The Role of Task Vectors
Laura Gomezjurado Gonzalez (equal contribution), Hiroki Naganuma (equal contribution), Kotaro Yoshida (equal contribution), Takafumi Horie, Yuji Naraki, Ryotaro Shimizu
International Conference on Learning Representations, January 2026
code
ICLR 2026
DiTaC: Conditioning Task Vectors via Distillation for Robust Model Merging
Kotaro Yoshida, Takafumi Horie, Yuji Naraki, Ryotaro Shimizu, Ioannis Mitliagkas, Hiroki Naganuma
International Conference on Learning Representations, January 2026
code
ICLR 2025
Mastering Task Arithmetic: τJp as a Key Indicator for Weight Disentanglement
Kotaro Yoshida, Yuji Naraki, Takafumi Horie, Ryosuke Yamaki, Ryotaro Shimizu, Yuki Saito, Julian McAuley, Hiroki Naganuma
International Conference on Learning Representations, April 2025
code travel grant
ICML 2024
No Wrong Turns: The Simple Geometry Of Neural Networks Optimization Paths
Charles Guille-Escuret (equal contribution), Hiroki Naganuma (equal contribution), Kilian Fatras, Ioannis Mitliagkas
International Conference on Machine Learning, July 2024
code
IJCNN 2023
How Image Corruption and Perturbation Affect Out-Of-Distribution Generalization and Calibration
Keigo Tada, Hiroki Naganuma
International Joint Conference on Neural Networks, June 2023
travel grant
AISTATS 2023
Conjugate Gradient Method for Generative Adversarial Networks
Hiroki Naganuma, Hideaki Iiduka
International Conference on Artificial Intelligence and Statistics, May 2023
poster code
CoLLAs 2022
Optimal Transport Meets Noisy Label Robust Loss and MixUp Regularization for Domain Adaptation
Kilian Fatras, Hiroki Naganuma, Ioannis Mitliagkas
Conference on Lifelong Learning Agents, August 2022
HPC Asia 2018
Accelerating Convolutional Neural Networks Using Low Precision Arithmetic
Hiroki Naganuma, Rio Yokota
International Conference on High Performance Computing in Asia-Pacific Region, January 2018
poster
HPCS 2017
Accelerating Matrix Multiplication in Deep Learning by using Low-Rank Approximation
Kazuki Osawa, Akira Sekiya, Hiroki Naganuma, Rio Yokota
International Conference on High Performance Computing & Simulation, July 2017

Preprints and under review(7)

arXiv
Understanding Stability of Outer Nesterov Momentum on Stochastic Pseudo-Gradients
Nishant Sharma, Laura Gomezjurado Gonzalez, Ganesh Talluri, Sarath Chandar, Hiroki Naganuma
arXiv, October 2026
pdf
arXiv
Which Geometry on Which Layer? Measured Sign-Spectral Assignment in Mixed-Optimizer Training
Hiroki Naganuma (equal contribution), Laura Gomezjurado (equal contribution), Mahdi Ghaznavi, Ganesh Talluri, Atsushi Nitanda, Seng Pei Liew, Ryuichiro Hataya, Ioannis Mitliagkas
arXiv, October 2026
pdf
Under Review
Orth-Dion: Eliminating Geometric Mismatch in Distributed Low-Rank Spectral Optimization
Tatsuhiro Nakamori (equal contribution), Laura Gomezjurado Gonzalez (equal contribution), Ganesh Talluri (equal contribution), Ansh Tiwari, Hideyuki Kawashima, Guillaume Rabusseau, Ioannis Mitliagkas, Hiroki Naganuma
Under Review, May 2026 (under review)
Under Review
Convergence Bound and Critical Batch Size of Muon Optimizer
Naoki Sato (equal contribution), Hiroki Naganuma (equal contribution), Hideaki Iiduka
Under Review, July 2025 (under review)
SSRN
When Does Alignment Help? A Comparative Study of DCCA and Fusion-Based Approaches for Multi-modal Chest X-ray Analysis
Hiroki Naganuma, Akira Uehara, Takeru Sakurai, Yoshiyuki Sankai
SSRN, April 2025
arXiv
Augmenting NER Datasets with LLMs: Towards Automated and Refined Annotation
Yuji Naraki (equal contribution), Ryosuke Yamaki (equal contribution), Yoshikazu Ikeda, Takafumi Horie, Kotaro Yoshida, Ryotaro Shimizu, Hiroki Naganuma
arXiv, March 2024
arXiv
Takeuchi's Information Criteria as Generalization Measures for DNNs Close to NTK Regime
Hiroki Naganuma, Taiji Suzuki, Rio Yokota, Masahiro Nomura, Kohta Ishikawa, Ikuro Sato
arXiv, September 2021

Workshop papers(9)

ICLR 2026 Workshop
The Geometry of Spectral Gradient Descent: Layerwise Criteria for SignSGD vs SpecSGD
Hiroki Naganuma (equal contribution), Laura Gomezjurado Gonzalez (equal contribution), Mahdi Ghaznavi (equal contribution), Ioannis Mitliagkas
ICLR 2026 Workshop on Geometry-grounded Representation Learning and Generative Modeling, February 2026
ICLR 2024 Workshop
Smoothness-Adaptive Sharpness Aware Minimization for Finding Flatter Minima
Hiroki Naganuma (equal contribution), Junhyung Lyle Kim (equal contribution), Anastasios Kyrillidis, Ioannis Mitliagkas
ICLR 2024 Workshop on Practical ML for Low Resource Settings, May 2024
workshop poster
NarSUM 2023
Story-to-Images Translation: Leveraging Diffusion Models and Large Language Models for Sequence Image Generation
Haruka Kumagai, Ryosuke Yamaki, Hiroki Naganuma
ACM Multimedia 2023 Workshop on User-centric Narrative Summarization of Long Videos, October 2023
workshop travel grant
UNCV 2023
On the Interplay of Curvature, Calibration and Out-of-Distribution Generalization: Insights from SAM and Focal Loss Analyses
Hiroki Naganuma (equal contribution), Masanari Kimura (equal contribution)
ICCV 2023 Workshop on Uncertainty Quantification for Computer Vision, October 2023
poster
ICLR 2023 Workshop
Necessary and Sufficient Hypothesis of Curvature: Understanding Connection Between Out-of-Distribution Generalization and Calibration
Hiroki Naganuma (equal contribution), Masanari Kimura (equal contribution)
ICLR 2023 Workshop on Domain Generalization, May 2023
MLSS 2020
Towards Understanding the Relationship of Batch Size and Iterations in Deep Learning
Hiroki Naganuma
Machine Learning Summer School 2020, June 2020
ACML 2019 Workshop
On Empirical Analysis of Layer-wise Learning Rate Schedule
Hiroki Naganuma, Rio Yokota
Asian Conference on Machine Learning 2019 Workshop, November 2019
HPML 2019
A Performance Improvement Approach for Second-Order Optimization in Large Mini-batch Training
Hiroki Naganuma, Rio Yokota
2nd High Performance Machine Learning Workshop, May 2019
slides
Tokyo Tech-Stony Brook Joint Meeting
Noise Injection Leads to Better Generalization in Large Mini-Batch Training
Hiroki Naganuma, Rio Yokota
Tokyo Institute of Technology and Stony Brook University Joint Science and Technology Meeting, May 2019

Equal contribution

Journal articles

Conference papers

On Empirical Analysis of Layer-wise Learning Rate Schedule

Noise Injection Leads to Better Generalization in Large Mini-Batch Training