Ryo Okumura
Physical AI Researcher at パナソニック コネクト
Berkeley, California, Japan
Summary
🤩
Rockstar🎓
Top SchoolRyo Okumura is a Physical AI Researcher based in Berkeley with 11 years of engineering experience bridging mechanical engineering and machine learning. Trained at the University of Tokyo with extended research at Stanford, he has progressed through technical leadership roles at Panasonic to his current research appointment at パナソニック コネクト. His work spans generative modeling and representation learning—evidenced by hands-on contributions integrating sparse attention and memory-efficient checkpointing into VideoGPT-style video generation pipelines. Ryo combines hardware-aware engineering instincts from a mechanical background with state-of-the-art ML methods to optimize model performance and efficiency. Known for pragmatic implementations that bring research ideas closer to production, he often focuses on attention sparsity and compute-aware model design. Colleagues value his rare mix of experimental rigor, cross-disciplinary perspective, and sustained delivery across industry research teams.
11 years of coding experience
修士（機械工学）, 情報理工学系研究科, 修士（機械工学）, 情報理工学系研究科 at 東京大学
Github Skills (9)
attention-mechanism10
video-production10
video-editing10
machine-learning10
pytorch10
video-encoder10
video-conversion10
deepspeed9
computer-vision9
Programming languages (6)
Github contributions (5)
wilson1yan/VideoGPT
Apr 2021 - Sep 2022
Role in this project:
ML Engineer
ML EngineerContributions:60 commits, 4 PRs, 62 pushes in 1 year 5 months
Contributions summary:Ryo's contributions focused on integrating and implementing sparse attention mechanisms within the VideoGPT framework. They added a `SparseAttention` module, including configurations for strided sparsity and attention masks, demonstrating a deep understanding of attention mechanisms. Furthermore, they integrated checkpointing within the transformer block and included necessary dependencies like `deepspeed.ops.sparse_attention`, indicating a focus on optimizing the model for performance and efficiency. This work directly supports the video generation capabilities of the project.
wilson1yan/wilson1yan.github.io
Nov 2019 - Jun 2024
Contributions:65 pushes in 4 years 7 months
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.