Yutaro Oguri

Yutaro Oguri

Yutaro Oguri

Undergraduate at The University of Tokyo, studying Computer Science and Engineering

東京大学 工学部 電子情報工学科

My research areas include approximate nearest neighbor search and multimedia information retrieval.
for large-scale data and its acceleration.
Experienced other engineering fields such as quantization of DNN,
backend development, distributed system, Kubernetes, and MLOps. Also working on some OSS projects.

機械学習とコンピュータビジョン、特に大規模データに対する検索や、高速化を研究しています。
その他、MLOps, Kubernetes, DBMSにも興味があります。OSS活動にも取り組んでいます。

Education

Education Icon

Bachelor of Engineering in Information and Communication Engineering (Apr. 2019 - Mar. 2024 (expected))

工学部 電子情報工学科 (Apr. 2019 - Mar. 2024 (卒業予定))

The University of Tokyo, Tokyo

東京大学

Advisor : Lecturer (Assistant Professor) Yusuke Matsui, Prof. Kiyoharu Aizawa

指導教員 : 松井 勇佑 講師, 相澤 清晴 教授

Study : Approximate Nearest Neighbor Search, Efficient Multimedia Retrieval

Study : Approximate Nearest Neighbor Search, Efficient Multimedia Retrieval

GPA : 3.88 / 4.30

GPA : 3.88 / 4.30

Job Experience

Experience Icon

Software Engineer (Part-time) (Mar. 2023 - Current)

M3, Inc., Tokyo, Part-time

エムスリー株式会社, Tokyo, Part-time

Working on the development of kannon, a new OSS library for parallel and distributed execution of gokart,
which is an OSS for machine learning pipelines, on Kubernetes.
I wrote an tech blog post introducing kannon library in Japanese.

機械学習パイプラインのOSSであるgokartを、Kubernetes上で並列分散実行するための新しいOSSライブラリであるkannonの開発に取り組んでいます。
kannonの技術詳細を説明するブログ記事を書きました。

Experience Icon

Research & Development Internship (Aug. 2022 - Sep. 2022)

Preferred Networks Inc., Tokyo, 2 months, Full-time

Preferred Networks Inc., Tokyo, 2 months, Full-time

I worked to improve the accuracy of chemical plant dynamics prediction by integrating deep learning and physical simulation.
I conducted an overall process from surveying previous research to examining the method, implementation, evaluation,
and presentation of the results.
The output cannot be disclosed due to confidentiality.

The specific technologies used are as follows.

  • Reproduction implementation of the paper on deep learning with PyTorch
  • The original implementation of Forward/Backward in PyTorch
  • CUDA
  • Rust and its Python binding

深層学習と物理シミュレーションを融合させることによる、化学プラントの挙動予測の精度向上に取り組みました。
先行研究の調査から、手法の検討、実装、評価、成果発表までを一貫して行いました。
成果は機密により公開できません

具体的な使用技術は次のとおりです.

  • 深層学習に関する論文の再現実装
  • PyTorchでのForward/Backwardの独自実装
  • CUDA
  • Rust and its Python binding

Experience Icon

Software Engineer (Part-time) (Mar. 2022 - Mar. 2023)

FLYWHEEL Inc., Tokyo, 1 year, Part-time

FLYWHEEL Inc., Tokyo, 1 year, Part-time

Worked on backend development of search and recommendation systems using Kotlin and Python.
In particular, I was mainly involved in query analysis, API specification, implementation, and load testing of the search engine.
Elasticsearch was used to implement the search engine.

Kotlin, Pythonを用いた検索・推薦システムのバックエンド開発に取り組みました。
特に検索エンジンのクエリ解析、API仕様策定、実装、負荷試験に携わりました。
検索エンジンの実装にはElasticsearchを使用しました。

Experience Icon

Software Engineer Internship (Jul. 2021 - Aug. 2021)

Fixstars Corporation, Tokyo, 1 month, Full-time

Fixstars Corporation, Tokyo, 1 month, Full-time

I worked on speeding up an open-sourced implementation of 3D reconstruction algorithms using parallel computation on NVIDIA GPUs,
achieving a speedup of about 11 times faster than the CPU implementation using CUDA.
The entire process, from catch-up to CUDA, bottleneck analysis using a profiler, implementation, and evaluation,
was done in 15 days.
The detailed output cannot be disclosed due to confidentiality.

GPUによる並列計算を用いた、3次元再構成アルゴリズムの高速化に取り組みました。
CPU実装に比べて約11倍の高速化を達成しました。
CUDAへのcatch upからプロファイラを用いたボトルネック解析、実装、評価までを15日で一貫して行いました。

Experience Icon

Software Engineer (Part-time) (Apr. 2021 - Mar. 2022)

Tier IV, Inc, Tokyo, 1 year Part-time

Tier IV, Inc, Tokyo, 1 year, Part-time

I worked on the implementation of a pipeline for autonomous driving data in Python, especially for point cloud data.
Also surveyed papers on object detection and worked on model acceleration using ONNX and TensorRT.

自動運転データ(特に点群)のためのパイプラインをPythonでの実装に取り組みました。
また、物体検出に関する論文のサーベイやONNX, TensorRTを用いたモデルの高速化に取り組みました。

Publication (referred)

Research Icon

Self-Examination Mechanism: Improving Standard Accuracy in Adversarial Purification

Sora SUEGAMI†, Yutaro OGURI†, Zaiying ZHAO†, Yu KAGAYA†, Koki MUKAI, Shun YOSHIDA, Fu CHEN, & Toshihiko YAMASAKI (2023).
"Self-Examination Mechanism: Improving Standard Accuracy in Adversarial Purification"
International Conference on Image, Video and Signal Processing (IVSP 2024). (†: equal contribution)

Sora SUEGAMI†, Yutaro OGURI†, Zaiying ZHAO†, Yu KAGAYA†, Koki MUKAI, Shun YOSHIDA, Fu CHEN, & Toshihiko YAMASAKI (2023).
"Self-Examination Mechanism: Improving Standard Accuracy in Adversarial Purification"
International Conference on Image, Video and Signal Processing (IVSP 2024). (†: equal contribution)

Research Icon

General and Practical Tuning Method for Off-the-Shelf Graph-Based Index

Yutaro OGURI, & Yusuke MATSUI (2023).
"General and Practical Tuning Method for Off-the-Shelf Graph-Based Index: SISAP Indexing Challenge Report by Team UTokyo"
International Conference on Similarity Search and Applications (SISAP 2023), 2023

Yutaro OGURI, & Yusuke MATSUI (2023).
"General and Practical Tuning Method for Off-the-Shelf Graph-Based Index: SISAP Indexing Challenge Report by Team UTokyo"
International Conference on Similarity Search and Applications (SISAP 2023), 2023

Publication (non-referred)

Research Icon

Self-Examination Mechanism: Lightweight Defense Mechanism against Adversarial Examples using Explainable AI.

Sora SUEGAMI†, Yutaro OGURI†, Zaiying ZHAO†, Yu KAGAYA†, Koki MUKAI, Shun YOSHIDA, Fu CHEN, & Toshihiko YAMASAKI (2023).
"Self-Examination Mechanism: Lightweight Defense Mechanism against Adversarial Examples using Explainable AI."
Proceedings of the Annual Conference of JSAI, JSAI2023, 2A1GS203-2A1GS203. (†: equal contribution)

末神 奏宙†, 小栗 悠太郎†, 趙 在瀛†, 加賀谷 湧†, 向井 皇喜, 吉田 舜, 琛 付, & 山崎 俊彦 (2023). Self-Examination Mechanism: 説明可能AIを用いた敵対的攻撃に対する軽量な防御機構. 人工知能学会全国大会論文集, JSAI2023, 2A1GS203-2A1GS203. (†: equal contribution)

Projects & Contributions

Works Icon

kannon

A library to execute data pipeline tasks on Kubernetes in a parallel and distributed manner with only minor changes.
I'm the original author of this library and currently maintaining it.
I wrote an detailed article in Japanese.
I presented about the library at k8s novice Tokyo #24 and MLOps study group.

gokartのTaskをKubernetes上で並列・分散実行するためのライブラリです。ほんの少しのコードの変更だけで簡単にk8s上で並列分散化することができます。
日本語で詳細な解説記事を書きました。
k8s novice Tokyo #24 にて発表しました。
【Sansan×エムスリー】gokartで爆速開発!MLOps勉強会にて発表しました。

Works Icon

Self-made Debugger using ptrace

A simple debugger implemented in C using `ptrace` system call. It can attach to a process and display the state of register.
I wrote a detailed explanation article in Japanese.

ptraceシステムコールを用いた簡易的なデバッガを自作しました。C言語で実装されています。Breakpointにおけるレジスタの中身を見ることができます。
日本語で詳細な解説記事を書きました。

Works Icon

More functional style CPython

I implemented a new feature to CPython's List to make it more functional style.
I wrote a detailed explanation article in Japanese.
This new feature enables to write code using Python's list like `list.map(func).filter(func).reduce(func)`.
It is a functional style usage of Python's list.
Gained hands-on experience with CPython's source code and got familiar with concepts of reference counting.

CPythonのListをより関数型的に扱うための新機能を実装しました。
日本語で詳細な解説記事を書きました。

Works Icon

py2cpp transpiler

A transpiler that converts a Python function into a C++ function.

Works Icon

minc

A C compiler implemented in Rust that supports a subset of C.

Works Icon

Cupy: 1 Contribution

Contributed to CuPy, a NumPy-compatible array library accelerated by CUDA.

Works Icon

eRusticDB: a toy document DB in Rust

A toy document DB implemented in Rust. It supports Elasticsearch-like queries.

Awards

Silver Icon

2nd place at SISAP 2023 Indexing Challenge

TOEFL iBT 95点 (Mar. 2023)

Got 2nd place out of 7 teams at SISAP 2023 Indexing Challenge in multiple tracks,
which is a competition on Approximate Nearest Neighbor Search
held at 16th International Conference onSimilarity Search and Applications (SISAP 2023).
Here is a link to the, my blog post, and the paper on our solution, which is accepted to SISAP 2023.

Language

Japanese 🇯🇵 (Native)

English 🇬🇧 (CEFR C1 level)

日本語 🇯🇵(母語)

英語 🇬🇧 (CEFR C1 level)

Certifications & Scores

Work Icon

TOEFL iBT Score 95 (Mar. 2023)

TOEFL iBT 95点 (Mar. 2023)

Work Icon

AtCoder Cyan (Highest 1213), My Profile

AtCoder 水色 (Highest 1213), My Profile