Masaki Saito
Independent Researcher at Preferred Networks, Inc.
Tokyo, Japan
Summary
🤩
Rockstar🎓
Top SchoolMasaki Saito is an independent researcher and machine learning engineer based in Tokyo with 13 years of experience bridging academic research and production-grade deep learning. He holds a PhD in Informatics from Tohoku University and has applied his expertise at Preferred Networks after serving as a JSPS research fellow, focusing on computer vision and neural network loss functions. A significant open-source contributor, he implemented core sigmoid cross-entropy functionality and replicated softmax support in prominent projects like Chainer and CuPy, demonstrating both CPU and GPU proficiency. His work on illustration2vec shows a practical bent for extracting semantic features from images and improving model evaluation (e.g., F1-based thresholding). Now freelancing, he combines rigorous academic training with hands-on engineering to solve real-world vision problems and optimize core ML primitives.
13 years of coding experience
1 year of employment as a software developer
学士, 理学部物理学科, 学士, 理学部物理学科 at 東北大学博士, システム情報科学専攻, 博士, システム情報科学専攻 at 東北大学大学院情報科学研究科
Japanese, English
Github Skills (17)
loss-functions10
python10
chainer10
testing10
gpu-programming10
machine-learning10
numpy10
deep-learning10
cupy10
neural-network10
cuda10
computer-vision10
softmax10
caffe9
cudnn9
Programming languages (1)
Github contributions (5)
rezoo/illustration2vec
Sep 2015 - Dec 2018
A simple deep learning library for estimating a set of tags and extracting semantic feature vectors from given illustrations.
Role in this project:
ML Engineer
ML EngineerContributions:1 release, 35 commits, 2 PRs in 3 years 3 months
Contributions summary:Masaki primarily contributed to the development and maintenance of an illustration-based deep learning library. They implemented methods for extracting semantic features from illustrations, including both feature extraction and binary feature generation. The user integrated Chainer framework support, expanded support for various image types, and incorporated options for thresholding based on F1 scores, demonstrating their focus on expanding the library's functionality and improving its performance.
pytorchdeep-learningvectorsestimatingdeep-learning-library
chainer/chainer
Jun 2015 - Sep 2018
A flexible framework of neural networks for deep learning
Role in this project:ML Engineer
ML EngineerContributions:273 commits, 118 PRs, 97 pushes in 3 years 3 months
Contributions summary:Masaki contributed significantly to the Chainer deep learning framework, adding and modifying functionalities related to neural network components. They implemented and tested a sigmoid cross-entropy function, essential for training models with sigmoid activations. Furthermore, the user worked on adding support for the replicated softmax layer, showing expertise in loss functions and their application within a deep learning context.
cudapythonmxnetcaffe2flexible-framework
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.