#Japan #Tokyo #AI Research #Asahi Shimbun #ICML 2026

Asahi Shimbun's Media Research Center Achieves Major Milestone at ICML 2026

The Asahi Shimbun Company, under the leadership of CEO Katsushi Tsunoda, proudly announces that its Media Research Development Center's research paper has been selected for presentation at the Forty-Third International Conference on Machine Learning (ICML 2026), one of the most prestigious international conferences in the field of machine learning. The paper, authored mainly by Shogo Yamauchi, aims at improving artificial intelligence (AI) models by proposing innovative methods that maintain accuracy while reducing computational time and costs.

Neural Networks (NN) serve as the foundation for AI models, and previous research has shown that using quaternion representations, known as Quaternion Neural Networks (Quaternion NN), can express complex models with fewer parameters than traditional real-number-based NNs. These networks have demonstrated compatibility particularly well with signal processing applications such as audio processing. In this groundbreaking study, improvements in the attention mechanism within Quaternion NNs were proposed, effectively decreasing the computational cost of both learning and inference processes. The result is a significant reduction of computational time while retaining predictability comparable to existing methodologies.

Attention Mechanism in AI

The attention mechanism is a critical component that directs focus towards important information, playing a vital role in the recent advancements in AI technologies. In the presented paper, comparisons were made of the new methods involving the quaternion NN's attention system against previous techniques, illustrating the enhanced efficiency and effectiveness of the new approach.

Asahi Shimbun is keen on applying the proposed techniques in the realm of audio and image signal processing for reporting purposes. Anticipated outcomes include faster transcription of interview audio and quicker analysis of large volumes of photographs and videos with reduced computational resources.

For those interested in a deep dive into this research, further details can be found on our technical blog: Asahi Shimbun's Media Technology Blog.

Details of the Research Paper

The full paper is titled Quaternion Self-Attention with Shared Scores, and was co-authored by Shogo Yamauchi, Tohru Nitta, and Hideaki Tamori. It is scheduled to be presented at ICML 2026, which will take place from July 6 to 11, 2026, in Seoul, South Korea. You can access the paper online at arXiv.

About ICML

The International Conference on Machine Learning, commonly referred to as ICML, is recognized as one of the leading global conferences in the machine learning domain. Each year, numerous submissions undergo a rigorous peer-review process, with only a fraction accepted, making it a distinguished platform for showcasing top-tier research. The acceptance of this paper reflects the high international esteem held for the research outcomes from the Media Research Development Center.

Overview of Media Research Development Center

Established in April 2021, the Media Research Development Center aims to leverage cutting-edge media technologies like artificial intelligence while utilizing resources inherent to newspaper operations, such as text, photographs, and audio. The Center focuses on addressing both internal and external organizational challenges through advanced methodologies, researching and developing techniques in areas like natural language processing and image processing.

For more insights into our AI research initiatives, visit Asahi Shimbun Media Research Development Center.