Movatterモバイル変換

Keisuke FukudaPreferred Networks, Inc.PFNにおける研究開発深層学習からMN-3開発，そして社員の働き方2022/10/19 融合情報学特別講義Ⅲ

自己紹介2● 福田圭祐 Keisuke Fukuda○ 東京工業大学(Tokyo Tech)○ Interests：■ High Performance Computing（HPC）● Perform large-scale parallel & distributed computing on supercomputers■ Joined PFN in Apr. 2017■ Distributed / Paralell Deep learning, performance optimization

Introduction toPreferred NetworksMaking the real world computable

Our Vision4We make cars, robots, and other devices more intelligent by fusing software and hardware in a sophisticatedmanner. By making devices intelligent enough to adapt to continuously changing environments and conditions, ourworld becomes computable through real-time sensing of the physical world.We do not compete in familiar territory, but rather take on ambitious technological challenges. By leveraging thelatest technologies, we want to advance the frontiers of knowledge and discover the world of the future.Making the real world computable.With our innovative and essential technologies,we venture into the unknown.

Company information5Manufacturing LogisticsTransportation Bio & HealthcarePersonal Robot EntertainmentFounded March 2014DirectorsCEO Toru NishikawaCER Daisuke OkanoharaCTO Ryosuke OkutaLocatedTokyo, Japan (HQ) Burlingame, CA., US(Preferred Networks America, Inc.)Number ofEmployees270+ Engineers & Researchers(October, 2020)

2021, 2020● No.1 on Green500 list of the world’smost energy-efficient supercomputers2019● Prime Minister’s Award, Nippon Venture Awards2018● Grand Prize37th NIKKEI Product and Service Excellence Award● Open Source Data Science Project Award, ODSC East 20182017● Japan-U.S. Innovation Awards「Emerging Leader Award」● FT ArcelorMittal Boldness in Business Awards● METI Minister’s Award, Nippon Venture Award2016● 1st Annual JEITA Venture Awards● Forbes JAPAN’s CEO OF THE YEAR 2016● 「1st place - Most innovative startup」Awards6

We develop practicalapplications of cutting-edgetechnologiesPreferred Networks (PFN) develops practicalapplications of deep learning and other cutting-edge technologies in order to solve real-worldproblems that are difficult to address with existingtechnologies.Our Focus7

Our Capabilities8Deep LearningWorld class researchersfocusing on deep learningExpertiseWide range of deep expertise fromrobotics to genomics tocomputational chemistryWorld class computationalresources designed for deeplearning applicationPrivateSuperComputerSoftwareIn-house developments of OSS andhyperparameter tuning library toaccelerate software development

● PFN collaborates with world-leading corporations and organizations to drive innovation in a wide range offields. We aim to build long-term relationships with our partners to create new innovations that lead tocreation of new businessesOur BusinessCutting-edge technologyxComputational resourcesBusiness challengesxHigh quality dataCreation ofnew businessesSoftware ApplicationsxIntellectual PropertyR&DprojectsProfitsharingPartneringCompany

Our Values10Preferred Networks is a young, yetrapidly growing companyOur Values are what make us differentAs PFN members, we question:what should we do and not do?who are we and what do we consider important?To answer these questions, we came up with thefour statements as our code of conduct, or PFN Values

● Employees: 300+ (270+ engineers & researchers)● Top Management + Corporate Officers = 12● Each team consists of an Engineering Manager + members○ Many members belong to multiple teams concurrently○ Slack-based communications, most channels are open● Working style under COVID-19 era○ WFH (Work-from-home) by default○ Slack / Zoom / Google Meet / Jamboard / Mural● We are exploring a new workign style for post COVID-19 eraHow we work in PFN11Teams (EM + 3-10 ppl. each)Corporate Officers (9)Top Management (3)

Industry automation powered by deep learning technologies13Autonomous learning for bin-picking robot.The robot gathers data by trial and error andlearns the place where it is likely to pick thepiece up by using deep learning (as ofDecember 2015).https://youtu.be/ydh_AdWZflA

@ICRA 2017 voice Recognition + object picking14“Interactively Picking Real-World Objects with Unconstrained Spoken Language Instructions”arXiv:1710.06280• ICRA is a top-tier conference on robotics• Best Paper Award on Human-Robot Interaction• Technologies:• Visual recognition• Natural language processing (NLP)• The robot can understand ambiguous words:• ”The Teddy bear”• ”The brown fluffy stuff”https://youtu.be/_Uyv1XIUqhk

Factory and plant operation control using deep learning15PFN is working with ENEOS (formerly JXTGNippon Oil & Energy Corporation) on a jointresearch project regarding optimization andautomation of oil refineries.An oil refinery is a very complex system consistingof hundreds of processes and thousands of sensorsand actuators.Because of its massive production scale ofpetrochemical products, an improvement of afraction of a percent of productivity deliverssignificant cost reductions.By leveraging PFN’s deep learning technology, thejoint venture aims to automatically control andoptimize large and complex plant equipment formore efficient use of energy resources.ENEOS’ Kawasaki Refinery. ML-based control model can be used to keep plant machine operationstable against unknown external disturbanceDL-based Digital Twins for advanced automation and optimization

17https://petalica-paint.pixiv.dev/index_ja.htmlロボット系エンジニアが、サイドプロジェクトとして開始→正式プロジェクトへ

Crypko™: High-quality Anime Character Generation and Design18Deep learning canrevolutionize theentertainment industryPFN’s technology Crypko uses state-of-the-art generative models, a branchof techniques in deep learning, togenerate a potentially infinite set ofunique, high quality characters notcontained in the training data.Furthermore, it can fuse severalcharacters into new characters,inheriting their distinctive featuresCrypko’s character fusion. From the two characters on the top row, Crypko can generate characters on the bottomrow that inherit distinctive features of the input characters For more information please visit our entertainment page:https://preferred.jp/en/projects/entertainment/

Playgram™ / Playgram typing™: Programming education for kids19Virtual, high-quality learning experience in Computer SciencePFN has developed Playgram™, a programming education app primarily targeting students in elementaryschool and above. PFN has teamed up with Yaruki Switch Group (YSG), Japan’s leading education group with adiverse range of programs and over 1,700 schools, to build a programming course package using Playgram.Beginning August 2020, YSG will first pilot the package in three schools in the Tokyo area, both in classrooms and online.Developed by PFN’s software engineers at the forefront of artificial intelligence technologies, Playgram incorporates the K-12 ComputerScience Framework, a U.S. guideline for computer science education. The app will be available in Japanese at launchFor more infromation, please visit our Playgram website: https://playgram.jp/Bridges the gapbetween visual andtext-based codingRich 3D interface thatinspires creativityAdaptive learningsystem and user-friendly tutorials

Optuna™: Automation for Hyper-parameter Tuning20Optimize Your OptimizationAn open source hyperparameter optimization framework to automate hyperparameter searchIn deep learning, it is essential to tune hyperparameters since they control how an algorithm behaves.The precision of a model largely depends on tuning a large number of hyperparameters, including training iterations, neuralnetwork layers and channels, learning rate, batch size, and so on. Optuna, an open source technology developed at PFN,automates this trial-and-error process of optimizing the hyperparameters. It automatically finds optimal hyperparametervalues that enable the algorithm to give excellent performance. Optuna can be used for any black-box optimization problems.

Consistent contributions to Research21Examples of recent PFN publications in top-tier conferences (in 2022)● [J. Chem. Inf. Model.2022] “Molecular Design Method Using a Reversible Tree Representation of ChemicalCompounds and Deep Reinforcement Learning”● [BMVC2022] “Multi-View Neural Surface Reconstruction with Structured Light”● [NeurIPS 2022] “Unsupervised Learning of Equivariant Structure from Sequences”● [NeurIPS 2022] “Decomposing NeRF for Editing via Feature Field Distillation”● [Lung Cancer 2022] “Machine Learning-based Exceptional Response Prediction of Nivolumab Monotherapywith Circulating MicroRNAs in Non-Small Cell Lung Cancer”● [ICAIF 2022] “Theoretically Motivated Data Augmentation and Regularization for Portfolio Construction”● [ICAIF 2022] “Efficient Learning of Nested Deep Hedging using Multiple Options”● [ROMAN 2022] “F3 Hand: A Versatile Robot Hand Inspired by Human Thumb and Index Fingers”● [Nature Communications 2022] “Towards universal neural network potential for material discovery applicableto arbitrary combination of 45 elements”● [Physical Review Research 2022] “Power Laws and Symmetries in a Minimal Model of Financial MarketEconomy”

Total of 2,560 GPUsTotal 200 PFLOPSListed No.1 in Japan amongst private entity1 PETA FLOPS =1,000 trillionFloating-point OperationsPer SecondOur Infrastructure23MN-Core MN-Core Board x 4CPU Intel Xeon 8260M 2way （48 physical cores）Memory 384GB DDR4Storage Class Memory 3TB Intel Optane DC Persistent MemoryNetworkMN-Core DirectConnect（112Gbps） x 2Mellanox ConnectX-6（100GbE） x 2On board（10GbE） x 2MN-3 specsDeep learning processor MN-CoreSupercomputer designed for deep learning applicationMN-1 MN-2 MN-3For more information please visit: https://projects.preferred.jp/supercomputers/en/

The MN series: PFN’s in-house supercomputers24 MN-1a (Sep. ’17〜)━ 1024 NVIDIA Tesla P100 + IB FDR━ Peak 19.1 Peta FLOPS (SP)━ #227 in Top500 Nov. 2018 MN-1b (July. ’18〜)━ 512 NVIDIA Tesla V100 + IB EDR━ Peak 57.3 Peta (tensor) Flops MN-2b (July. ’19〜)━ 1024 NVIDIA Tesla V100 + IB EDR━ 128 Peta (Tensor) Flops MN-3 (Nov. 20〜)━ We’ll later!

010203040506070Time[min]Training time of ResNet-50 (90 epochs) on ImageNetAchievement on MN-1a: ImageNet in 15 minutes252018 July2018 Nov2017 NovarXiv: 1711.04325Extremely Large Minibatch SGD: Training ResNet-50on ImageNet in 15 Minutes2018 Nov

Achievement on MN-1b: PFDet in OIC 201926

● Google AI Open Images - Object Detection Track○ Competition using Largest-class image dataset○ 12 million bounding boxes, 1.7 million images○ 454 competitiors○ Approx. 500GB (annotated subset)● Object detection: much harder than object recognition taskAchievement on MN-1b: PFDet in OIC 201827

https://tech.nikkeibp.co.jp/atcl/nxt/column/18/01006/101000005/28

○ c.f. 「演繹から帰納へ〜新しいシステム開発パラダイム〜」丸山宏, PPL2018 招待講演○ 特別なものではなく、実装手法の1つとして広く使われるようになっていくのでは？AIはコンピューターサイエンスのコア技術になっていくAIが向いている場面 AIが不向きな場面• データが大量（or 生成可能）• 誤差が許容される• 現象が複雑／原理が不明• シミュレーションが困難／計算量多い• 法則・原理が一定• 予測が目的• データが少ない• 厳密さが必要• 演繹的プログラミングが可能• シミュレーションが容易／手法が確立• 過去から未来が予測できない• メカニズムの理解が目的⇒困難なタスクは計算パワーで解く⇒計算パワーが無いと戦えないConventional Programming従来のプログラミング演繹的プログラミング(Deductive programming)Machine Learning機械学習帰納的プログラミング(Reductive programming)

Simulationとは：● 現実世界の物理法則を数式でモデル化し、計算機上で計算によって再現・予測する● 流体、天体、気象、機械設計、材料化学、・・・Simulationの課題● 複雑すぎる現象・Multiphysics（ex. 構造連成計算、気象）● 計算量の爆発Simulation31

● これまで深層学習の実用化はデータが容易に入手可能な分野（ウェブ、バーチャル）に限られていた。● 今後、実世界の問題に深層学習を導入していくためにはシミュレーション利用が不可欠である● データが21世紀の石油と言われる中で、そのデータ自身を作れるシミュレーションを揃えていくことが重要となる● またシミュレーション自体も深層学習を利用することで劇的に高速化、多様化を達成できる今後シミュレーションが重要となる32

SimulationとAI は相性が良い33Simulationの中でも難しいとされているものに対して、AIを組み合わせて互いに補い合うAIが向いている場面 AIが不向きな場面• データが大量（or 生成可能）• 誤差が許容される• 現象が複雑／原理が不明• シミュレーションが困難／計算量多い• 法則・原理が一定• 予測が目的• データが少ない• 厳密さが必要• 演繹的プログラミングが可能• シミュレーションが容易／手法が確立• 過去から未来が予測できない• メカニズムの理解が目的Simulationが向いている場面 Simulationが不向きな場面• 少ない物理法則から、モデル化可能• 保存則などを厳密に維持可能• メカニズムの理解・予測の両方• 現象が複雑・原理が不明なものは難• 計算量が爆発するSimulationが深層学習を助ける• 網羅的なデータを入手可能• ラベルを作るのが難しい場合もラベル付が可能• 最適化、強化学習に必要なWhat-If分析が可能深層学習がSimulationを助ける• シミュレーションの高速化• データからシミュレーションを学習する• データ同化、パラメータ推定を助ける

34https://matlantis.comAI x Simulationの事例（１）

● 2022年度夏季インターンシップの成果（東京大学・助田さん）● 気象シミュレーションは、シミュレーションの中でも特に難しい分野○ 観測データが少ない（観測機器の制約）○ 計算量が多い○ 現象が複雑● このテーマでは、「計算量が多い」という課題に着目して、スパコンで実行されるシミュレーターを省メモリで模倣計算することにチャレンジ数値シミュレーションデータの低次元潜在空間における時間発展ダイナミクスの学習35AI x Simulationの事例（２）Preferred Networks Tech Blog “数値シミュレーションデータの低次元潜在空間における時間発展ダイナミクスの学習”

このあと続くDeep Learningのための専用プロセッサ「MN-Core」の開発と活用金子紘也 Hiroya Kaneko● PFNにとっての計算能力の位置付け● 代表的なDeep Learningの高速化手法● なぜ今プロセッサ開発なのか？● MN-Coreの概要● 開発チームの働き方● 最近の成果

Movatterモバイル変換

Change Language

PFNにおける研究開発（2022/10/19 東大大学院「融合情報学特別講義Ⅲ」）

In this document

Recommended

More Related Content

What's hot

Similar to PFNにおける研究開発（2022/10/19 東大大学院「融合情報学特別講義Ⅲ」）

More from Preferred Networks

Recently uploaded

PFNにおける研究開発（2022/10/19 東大大学院「融合情報学特別講義Ⅲ」）