Yoshinari Takeishi (武石 啓成)

Journal Papers

1. Y. Takeishi, M. Iida & J. Takeuchi:
   “Approximate Spectral Decomposition of Fisher Information Matrix for Simple ReLU Networks,” Neural Networks, Vol. 164, pp. 691-706, July 2023. (Link)
2. Y. Takeishi & J. Takeuchi:
   “An Improved Analysis of Least Squares Superposition Codes with Bernoulli Dictionary,” Japanese Journal of Statistics and Data Science, 2, pp. 591-613, September 2019.
3. Y. Takeishi, M. Kawakita, & J. Takeuchi:
   “Least Squares Superposition Codes with Bernoulli Dictionary are Still Reliable at Rates up to Capacity,” IEEE Transactions on Information Theory, Vol. 60, No. 5, pp. 2737-2750, May 2014.

Conference Papers (peer reviewed)

1. Y. Takeishi, M. Iida, S. Wang, & J. Takeuchi:
   “A Novel Stabilizing Technique to Accelerate Stochastic Gradient Descent in Simple ReLU Networks,” Proc. of 2026 International Symposium on Information Theory and Its Applications (ISITA), pp. 634-639, Okinawa, Japan, November 1-4, 2026.
2. Kazuhiro Aki, Yoshinari Takeishi, Yutaka Jitsumatsu, Shigehide Kuhara, Jun'ichi Takeuchi, & Hidenori Takeshima:
   “Synthesis of Three-dimensional (3D) Multi-contrast Brain Tumor Images with Controlled Tumor Properties Using Generative Models,” Proc. Intl. Soc. Mag. Reson. Med, 2025.
3. Y. Takeishi & J. Takeuchi:
   “Risk Bound on MDL Estimator for Simple ReLU Networks,” Proc. of 2024 IEEE International Symposium on Information Theory, pp. 256 - 261, Athens, Greece, July 7-12, 2024.
4. M. Iida, Y. Takeishi, & J. Takeuchi:
   “On Fisher Information Matrix for Simple Neural Networks With Softplus Activation,” Proc. of 2022 IEEE International Symposium on Information Theory, pp. 3014 - 3019, Espoo, Finland, June 26-July 1, 2022.
5. Y. Takeishi & J. Takeuchi:
   “An Improved Upper Bound on Block Error Probability of Least Squares Superposition Codes with Unbiased Bernoulli Dictionary,” Proc. of 2016 IEEE International Symposium on Information Theory, pp. 1168 - 1172, Barcelona, Spain, July 10-15, 2016.
6. Y. Takeishi, M. Kawakita, & J. Takeuchi:
   “Least Squares Superposition Codes with Bernoulli Dictionary are Still Reliable at Rates up to Capacity,” Proc. of 2013 IEEE International Symposium on Information Theory, pp. 1396-1400, Istanbul, Turkey, July 7-12, 2013.

Preprint

1. V. R. Takahashi, Y. Takeishi, J. Takeuchi, K. Salamatian:
   “Improving the Robustness/Accuracy Tradeoff Against Adversarial Attacks Using Information Bottleneck Distillation Through Dual Teachers,” arXiv:2607.27737, 2026.
2. J. Takeuchi, Y. Takeishi, N. Murata, K. Mimura, K. L. K. Ho, & H. Nagaoka:
   “Neural Tangent Kernels and Fisher Information Matrices for Simple ReLU Networks with Random Hidden Weights,” arXiv:2507.18555, 2025.
3. K. L. K. Ho, Y. Takeishi, & J. Takeuchi:
   “Function Forms of Simple ReLU Networks with Random Hidden Weights,” arXiv:2505.17907, 2025.
4. Y. Takeishi & J. Takeuchi:
   “Sparse Superposition Codes with Binomial Dictionary are Capacity-Achieving with Maximum Likelihood Decoding,” arXiv:2504.14262, 2025.
5. Y. Takeishi & J. Takeuchi:
   “Risk Bounds on MDL Estimators for Linear Regression Models with Application to Simple ReLU Neural Networks,” arXiv:2407.03854, 2024.
6. Y. Takeishi, M. Iida, & J. Takeuchi:
   “Approximate Spectral Decomposition of Fisher Information Matrix for Simple ReLU Networks,” arXiv:2111.15256, 2021.
7. Y. Takeishi & J. Takeuchi:
   “An Improved Analysis of Least Squares Superposition Codes with Bernoulli Dictionary,” arXiv:1801.02930, 2018.

2026

• 深層学習の理論基盤構築に向けたFisher情報行列の解析と応用, 日本応用数理学会2026年度年会, September 16, 2026.
• Toward Theoretical Foundations of Deep Learning: the MDL Principle and the Stabilizer Method, Mathematical Foundations of AI Seminar, Umeå University, Sweden, May 21, 2026.
• Towards a Theoretical Understanding of Deep Learning via the Minimum Description Length Principle, Information Theory Seminar, University of Cambridge, UK, February 4, 2026.
• The Minimum Description Length Principle and Machine Learning, Information Theory Seminar, University of Cambridge, UK, January 28, 2026.

2025

• Toward a Theoretical Foundation of Deep Learning: An MDL Perspective, MIAI Meeting, Université Grenoble Alpes, France, November 13, 2025.

2024

• MDL原理から見た深層学習の理論, 第47回情報理論とその応用シンポジウム (SITA2024) ワークショップ, December 11, 2024.

2023

• スパース重ね合わせ符号の最適性能と効率的復号, 離散数学とその応用研究集会2023, August 28, 2023.

2022

• A novel method to stabilize gradient descent learning, MIRAI2.0 Research & Innovation Week 2022, November 16, 2022. (Slides)
• On Sparse Superposition Codes and Discretization of Their Dictionaries, Mathematics for Innovation in Telecommunications Technology, September 15–16, 2022. (Slides)

• Information theory (Coding theory, MDL principle)
  スパース重ね合わせ符号についての 解説論文 を書きました(2023年7月)
• Machine learning theory (Neural networks)

Education

• 2022 Dr.Eng.
  Graduate School of Information Science and Electrical Engineering, Kyushu University, Japan.
  Thesis: Some Studies on Linear Regression Problems in Sparse Superposition Codes and Neural Networks.
Advisor: Prof. Jun'ichi Takeuchi.
• 2013 M.Eng.
  Graduate School of Information Science and Electrical Engineering, Kyushu University, Japan.
  Thesis: A Study on Discrete Sparse Superposition Codes for Additive White Gaussian Noise Channel.
  Advisor: Prof. Jun'ichi Takeuchi.
• 2011 B.Eng.
  School of Engineering, Kyushu University, Japan.

Job Experience

• April 2022--present.
  Assistant Professor, Faculty of Information Science and Electrical Engineering, Kyushu University.
• April 2013--February 2021.
  Server Engineer at Mitsubishi Electric Information Network Corporation.

takeishi(at)inf.kyushu-u.ac.jp