Asahi Shimbun's Media Research & Development Center Selected for ICML 2026, a Leading International Conference in Machine Learning

The Asahi Shimbun Company (President and CEO: Masaru Kadota) announces that a research paper from its Media Research & Development Center has been accepted at the 'Forty-Third International Conference on Machine Learning (ICML 2026),' one of the world's most prestigious international conferences in the field of machine learning. The paper, primarily authored by Shogo Yamauchi of the Media Research & Development Center, proposes a new method to reduce computational time and cost while preserving the accuracy of AI models.

Neural networks (NN) are a foundational technology for AI models. 'Quaternion Neural Networks (Quaternion NN),' which represent data using quaternions (※1), have been shown in prior studies to require fewer parameters than real-number-based NNs and are particularly effective for signal processing tasks such as speech. This study proposes an improved attention mechanism (※2) for Quaternion NNs, further reducing computational costs during training and inference. As a result, the method achieves a significant reduction in computation time while maintaining prediction performance comparable to existing approaches.

※1: Quaternions are four-dimensional extensions of complex numbers. For further explanation of quaternions and Quaternion NNs, please refer to the following blog post written by the researcher:

https://qiita.com/Yamachi-s/items/699256ab1b09b52dc98a

※2: The attention mechanism is a technique that enables AI models to focus on important information and is a key component in modern AI systems.

Comparison of attention mechanisms using Quaternion NNs (quoted from the paper)

The Asahi Shimbun Company plans to apply this proposed method to signal processing tasks such as audio and image analysis, aiming to utilize it in journalistic operations. For example, it is expected to enable faster transcription of interview recordings and more efficient analysis of large volumes of photos and videos using fewer computational resources.

Detailed explanations of the research are available on the center's technical blog:

note: https://note.com/asahi_ictrad/n/nde78eabe8457

About the Paper

Shogo Yamauchi, Tohru Nitta, Hideaki Tamori. Quaternion Self-Attention with Shared Scores. In Proceedings of the Forty-Third International Conference on Machine Learning (ICML 2026), Seoul, South Korea, July 2026.

https://arxiv.org/pdf/2605.24920

Japanese Title: Quaternion Self-Attention with Shared Scores

About ICML

ICML (International Conference on Machine Learning) is one of the world's leading academic conferences in machine learning. Papers are rigorously peer-reviewed, and only a select few are accepted from thousands of submissions worldwide. This acceptance demonstrates the international recognition of the Media Research & Development Center's research excellence. ICML 2026 will be held in Seoul, South Korea, from July 6 to 11, 2026.

About the Media Research & Development Center

Established in April 2021, the Media Research & Development Center leverages cutting-edge media technologies such as artificial intelligence, combined with the Asahi Shimbun's unique assets—including text, photos, and audio—to solve internal and external challenges. The center actively conducts research and development in advanced technologies such as natural language processing and image processing.

Asahi Shimbun Media Research & Development Center: AI Research Initiatives (https://cl.asahi.com/)