SlideShare a Scribd company logo
Azure Machine Learning
最新動向のご紹介
2020/06/24
日本マイクロソフト株式会社
女部田啓太
#dllab
konabuta
Keita Onabuta
Cloud Solution Architect
- Machine Learning & Deep Learning
• Data Scientist、機械学習エンジニア
• AI プロジェクト担当者・責任者
• 機械学習の導入のハードルの高さや運用管理の煩雑さに
課題を抱えている
• Azure Machine Learning の最新動向を 4 つの特徴の観点から
ご理解いただき、皆様の機械学習プロジェクトでご活用いただく
対象者
課題
ゴール
Azure Machine Learning
モデル学習 パッケージ化
& モデル検証
アセット管理 モニタリングデプロイ
CI/CD & モデル再学習
Azure DevOps & GitHub 統合
TrustedIndustry leading
MLOps
Open &
Interoperable
For all skill
levels
あらゆるスキルレベルに対応し、
ML の生産性・利便性を向上
DevOps 連携による
ML ライフサイクルの運用管理
責任のある
ML ソリューションの構築
オープンテクノロジーの採用
と相互運用性の実現
TrustedIndustry leading
MLOps
Open &
Interoperable
For all skill
levels
AutoML BERT
No-Code Deep Learning
GitHub Actions Responsible MLNotebook
TrustedIndustry leading
MLOps
Open &
Interoperable
GitHub Action Responsible MLNotebook
For all skill
levels
AutoML BERT
No-Code Deep Learning
参加はこちらから↓
自然言語処理ナイト
https://dllab.connpass.com/event/177785/
• 直感的なマウス操作でパイプライン構築
• 特徴量エンジニアリング
• モデル学習 (回帰、分類、クラスタリング)
•
• 推論 (リアルタイム & バッチ)
• カスタムモデル・スクリプト (Python, R)
機械学習パイプライン構築、テスト、デプロイするためのビジュアルワークフロー
※ 参考 : Azure Machine Learning デザイナー (プレビュー) とは
https://docs.microsoft.com/ja-jp/azure/machine-learning/concept-designer
Deep Learning による画像認識モデルを追加
• PyTorch を採用
• 基本的な画像の前処理
• 学習済みモデルによる転移学習
レコメンデーションのアルゴリズム強化
• wide & deep recommender
• 協調フィルタリングとコンテンツベースのハイブリッド型 Wide & Deep Learning for Recommender Systems” by Cheng et al.
※参考 : Algorithm & module reference for Azure Machine Learning designer (preview)
https://docs.microsoft.com/ja-jp/azure/machine-learning/algorithm-module-
reference/module-reference?view=azure-ml-py
Demo ~Designer~
・PyTorch の画像認識モデル
・レコメンデーションモデルの機能確認
Azure Machine Learning Build 2020
For all skill
levels
TrustedOpen &
Interoperable
Responsible MLNotebook
AutoML BERT
No-Code Deep Learning
Industry leading
MLOps
GitHub Actions
Data Scientist IT Professional
モデル学習の再現性 モデル再学習の自動化シンプルなデプロイモデル検証
"エンドーツーエンドの機械学習ライフサイクルを実現”
Data Scientist IT Professional
モデル学習の再現性 モデル再学習の自動化シンプルなデプロイモデル検証
"エンドーツーエンドの機械学習ライフサイクルを実現”
Azure DevOps
& Github
Azure Machine Learning
Github 上で Azure ML の CI/CD を構築するための
Github Actions を Github Marketplace にて公開
GitHub Actions Azure ML
• エンタープライズレベルの CI/CD によって
GitHub によるワークフローを自動化
• GitHub とネイティブ統合
• 3,400 以上の Actions が
Marketplace で公開
• マネージド or セフルホステッド環境が
利用可能
※参考 : GitHub Actionsについて
https://help.github.com/ja/actions/getting-
started-with-github-actions/about-github-actions
Workspace Compute Action Run Action
Register Model
Action
Deploy Action
Azure ML Workspace
の作成と接続
GPU/CPU サーバ
の作成と接続
Python スクリプトや
ML パイプラインの実行
学習済みモデルの登録
Kubernetes へ
モデルのデプロイメント
Demo ~MLOps~
git commit をトリガーにした CI/CD の実行
MLOps with GitHub Actions and Azure Machine Learning (デモで利用)
https://github.com/machine-learning-apps/ml-template-azure
Azure Machine Learning Build 2020
Industry leading
MLOps
GitHub Action
For all skill
levels
Trusted
Responsible MLAutoML BERT
No-Code Deep Learning
Open &
Interoperable
Notebook
nteract & Monaco Editor を UI に採用
コラボレーション機能 (coming soon)
Python スクリプトの自動生成 (coming soon)
Compute Instances をエンジン
• 普段使い慣れた Jupyter Notebook
& Jupyter Lab への切り替えも容易
https://channel9.msdn.com/Shows/AI-Show/Edit-and-run-Jupyter-notebooks-without-
leaving-Azure-Machine-Learning-studio
by Kaitlin McKinnon, Azure Machine Learning Program Manager
Industry leading
MLOps
GitHub Action
For all skill
levels
Open &
Interoperable
Notebook
AutoML BERT
No-Code Deep Learning
Trusted
Responsible ML
責任のある機械学習ライフサイクル
を実現するための包括的な機能
Understand Protect Control
モデルの理解と公平性の実現 プライバシー保護と機密性の確保 機械学習プロセス管理とガバナンス
Understand Protect Control
モデルの理解と公平性の実現 プライバシー保護と機密性の確保 機械学習プロセス管理とガバナンス
Fairlearn for
Classification, Regression
https://github.com/fairlearn/fairlearn
https://github.com/fairlearn/fairlearn
公平性の評価 :
公平性を評価する一般的なメトリックと
ダッシュボードを利用した Sensitive
Feature の評価
モデルのフォーマット :
scikit-learn, TensorFlow, PyTorch,
Keras などに対応
メトリック:
15以上の一般的なグループを対象にした公
平性メトリック
モデルの種類:
クラス分類、回帰
不公平性の軽減 :
最先端のアルゴリズムによって
分類・回帰モデルの不公平性を軽減
Demo ~Fairlearn~
ローン審査モデルにおける不公平性の Mitigate (軽減)
※ サンプルコード:Using Fairlearn GridSearch with census data
https://github.com/fairlearn/fairlearn/blob/master/notebooks/Grid%20Search%20
with%20Census%20Data.ipynb
Azure Machine Learning Build 2020
Azure Machine Learning Build 2020
Interpretability for
Tabular data
https://github.com/interpretml/interpret-
community
https://interpret.ml/
解釈可能性が高い構造を持つ
機械学習アルゴリズム
Fever?
Internal
Bleeding?Stay
Home
Stay
Home
Go to
Hospita
l
決定木
ルールリスト
線形回帰・ロジスティック回帰
Explainable Boosting Machines
….
Glass-box
models
ブラックボックスな
機械学習モデルの説明
Model
ExplanationPerturb
Inputs
Analyze
SHAP
LIME
Partial Dependence
Sensitivity Analysis
Black-box
explanations
各特徴量がどのくらい寄与
House price
prediction:
$300,000
Average house
price prediction
for all apartments
is $310,000
Delta here is
- $10,000
Parks contributed
+10k
Cats banned
contributed -20k
Secure neighborhood
contributed +10k
Wall painting needed
contributed -10k
各特徴量がどのくらい寄与
https://channel9.msdn.com/Shows/AI-Show/The-Science-Behind-InterpretML-SHAP
by Scott Lundberg, SHAP Founder @Microsoft Research
Demo ~InterpretML~
退職者予測モデルの解釈
※ 後日公開予定の最新ダッシュボードを利用
Azure Machine Learning Build 2020
TrustedIndustry leading
MLOps
Open &
Interoperable
For all skill
levels
AutoML BERT
Designer for Deep Learning
GitHub Actions Responsible MLNotebook
参考情報
Open Source Repo Link
Microsoft ML Tech List Microsoft の機械学習テクノロジー集 https://aka.ms/microsoft-ml-tech
Azure ML Notebook Azure Machine Learning 公式サンプルコード https://aka.ms/ml-notebooks
BERT Large 自然言語モデル BERT のサンプルコード http://aka.ms/azure-bert
LightGBM LightGBM トップページ https://aka.ms/lightgbm
ONNX ONNX トップページ https://aka.ms/onnx
Distributed DL 分散深層学習サンプルコード https://aka.ms/distdl
機械学習レシピ集 Link
Computer Vision コンピュータビジョン (Faster-RCNN ...) https://aka.ms/cv-recipes
NLP 自然言語処系 (BERT, GenSen ...) https://aka.ms/bert-recipes
Recommenders レコメンデーション (NCF, SAR ...) https://aka.ms/reco-recipes
Forecasting 時系列予測 https://aka.ms/forecast-recipes
MLOps MLOps CI/CD・コード集 https://aka.ms/cicd-recipes
Azure Machine Learning Build 2020
Azure Machine Learning Build 2020

More Related Content

PDF
データから価値を生み続けるには
PPTX
機械学習 - MNIST の次のステップ
PDF
[第37回 Machine Learning 15minutes!] Microsoft AI - Build 2019 Updates ~ Azure ...
PPTX
【Microsoft×Aidemy】初めて作る!機械学習モデルハンズオンセミナー: Designer の知っておきたい便利機能
PPTX
【Microsoft×Aidemy】初めて作る!機械学習モデルハンズオンセミナー - Azure Machine Learning 環境設定
PPTX
POWA50 - Power Apps と Power Automate の大規模な管理とサポート
PDF
Real-time personalized recommendation using embedding
PDF
マイクロソフトのPower Platformが描く、広島の世界
データから価値を生み続けるには
機械学習 - MNIST の次のステップ
[第37回 Machine Learning 15minutes!] Microsoft AI - Build 2019 Updates ~ Azure ...
【Microsoft×Aidemy】初めて作る!機械学習モデルハンズオンセミナー: Designer の知っておきたい便利機能
【Microsoft×Aidemy】初めて作る!機械学習モデルハンズオンセミナー - Azure Machine Learning 環境設定
POWA50 - Power Apps と Power Automate の大規模な管理とサポート
Real-time personalized recommendation using embedding
マイクロソフトのPower Platformが描く、広島の世界

What's hot (20)

PDF
なぜ今なのか?非開発者でも開発者でも、これから始めるPower Platform
PPTX
InterBEE 2020 Microsoft - AI for Media - Technical Deep Dive
PDF
Microsoft Power Platformで組織に力を与えよう
PDF
メルペイの与信モデリングで安全・安心のために実践していること
PPTX
Microsoft PowerApps ビジョンとロードマップ 2019年秋
PDF
SQLを書くだけでAPIが作れる基盤
PDF
Jupyter だけで機械学習を実サービス展開できる基盤
PPTX
stapy_028_talk1
PDF
AWSからOpenStack, Chef SoloからChef Serverに インフラを置き換えた事例の紹介
PPTX
Microsoft Power Platform がエンジニアにも必要な理由
PDF
アジャイル開発のためのDatadog
PPTX
Deep Learning Lab : Build 2020 Update - Reinforcement Learning on Azure Machi...
PDF
データアナリティクスの新たな一歩とそれを支えるインフラ
PDF
Decode19 cd42 fixer_public_0601
PDF
The Design for Serverless ETL Pipeline データ分析基盤のレガシーなデータロードをサーバレスでフルリプレースするまで道のり
PDF
Ibm data science experience
PDF
Awsで作るビッグデータ解析今とこれから
PPTX
Power Apps? なにそれ? おいしいの?
PPTX
Work Automate with Power Automate
PDF
Global Network on Alibaba Cloud
なぜ今なのか?非開発者でも開発者でも、これから始めるPower Platform
InterBEE 2020 Microsoft - AI for Media - Technical Deep Dive
Microsoft Power Platformで組織に力を与えよう
メルペイの与信モデリングで安全・安心のために実践していること
Microsoft PowerApps ビジョンとロードマップ 2019年秋
SQLを書くだけでAPIが作れる基盤
Jupyter だけで機械学習を実サービス展開できる基盤
stapy_028_talk1
AWSからOpenStack, Chef SoloからChef Serverに インフラを置き換えた事例の紹介
Microsoft Power Platform がエンジニアにも必要な理由
アジャイル開発のためのDatadog
Deep Learning Lab : Build 2020 Update - Reinforcement Learning on Azure Machi...
データアナリティクスの新たな一歩とそれを支えるインフラ
Decode19 cd42 fixer_public_0601
The Design for Serverless ETL Pipeline データ分析基盤のレガシーなデータロードをサーバレスでフルリプレースするまで道のり
Ibm data science experience
Awsで作るビッグデータ解析今とこれから
Power Apps? なにそれ? おいしいの?
Work Automate with Power Automate
Global Network on Alibaba Cloud
Ad

Similar to Azure Machine Learning Build 2020 (20)

PPTX
BrainPad - Doors - A-1 - Microsoft Data and AI
PDF
Azure machine learning service 最新の機械学習プラットフォーム
PDF
Azure Machine Learning アップデートセミナー 20191127
PPTX
Azure Machine Learning Hands-on - AutoML編 - 2020年1月
PDF
[Japan Tech summit 2017] MAI 003
PDF
AIビジネスクリエーションワークショップ@東京
PPTX
東北大学AIE - 機械学習中級編とAzure紹介
PPTX
Qiita x Microsoft - 機械学習セミナー Microsoft AI Platform
PDF
【de:code 2020】 アマダの Azure への取り組みと DevOPS・MLOPS 環境の構築と運用
PDF
[Microsoft Tech Summit 2018] Azure Machine Learning サービスと Azure Databricks で実...
PPTX
Try Azure Machine Learning
PPTX
Azure Machine Learning services 2019年6月版
PDF
[Developers Festa Sapporo 2018] Azure AI ~Microsoft AzureでのAI開発のイマ~
PDF
Edge から Cloud, Beginner から Professional までサポートする Azure AI プラットフォーム
PPTX
Azure Machine Learning Services 概要 - 2019年3月版
PPTX
機械学習 / Deep Learning 大全 (5) Tool編
PDF
Gpu accelerates aimodeldevelopmentandanalyticsutilizingelasticsearchandazure ai
PPTX
Microsoft Open Tech Night: Azure Machine Learning - AutoML徹底解説
PDF
講演資料「Azure AI Update Ignite Fall 2021を振り返ろう!」
PDF
【de:code 2020】 学生諸君! Azure であそぼう! ~Microsoft の AI を自作アプリに取り入れるまで~
BrainPad - Doors - A-1 - Microsoft Data and AI
Azure machine learning service 最新の機械学習プラットフォーム
Azure Machine Learning アップデートセミナー 20191127
Azure Machine Learning Hands-on - AutoML編 - 2020年1月
[Japan Tech summit 2017] MAI 003
AIビジネスクリエーションワークショップ@東京
東北大学AIE - 機械学習中級編とAzure紹介
Qiita x Microsoft - 機械学習セミナー Microsoft AI Platform
【de:code 2020】 アマダの Azure への取り組みと DevOPS・MLOPS 環境の構築と運用
[Microsoft Tech Summit 2018] Azure Machine Learning サービスと Azure Databricks で実...
Try Azure Machine Learning
Azure Machine Learning services 2019年6月版
[Developers Festa Sapporo 2018] Azure AI ~Microsoft AzureでのAI開発のイマ~
Edge から Cloud, Beginner から Professional までサポートする Azure AI プラットフォーム
Azure Machine Learning Services 概要 - 2019年3月版
機械学習 / Deep Learning 大全 (5) Tool編
Gpu accelerates aimodeldevelopmentandanalyticsutilizingelasticsearchandazure ai
Microsoft Open Tech Night: Azure Machine Learning - AutoML徹底解説
講演資料「Azure AI Update Ignite Fall 2021を振り返ろう!」
【de:code 2020】 学生諸君! Azure であそぼう! ~Microsoft の AI を自作アプリに取り入れるまで~
Ad

More from Keita Onabuta (8)

PDF
Azure における強化学習への取り組み
PDF
Azure Machine Learning NLP 最新動向 2020/07/02
PDF
DataDrift in Azure Machine Learning
PDF
Azure ML - OpenVINO 師走の推論ナイト 20191211 講演資料
PDF
AutoML & InterpretML (2019/11/27 Deep Learning Lab 講演資料)
PDF
品質管理 & 設備保全のための機械学習入門 2019年9月 (Quality Control and Predictive Maintenance App...
PDF
Neural Network Intelligence 概要 (AutoML Platform)
PDF
[DLLAB Engineer Days 講演資料] Automated ML & Model Interpretability
Azure における強化学習への取り組み
Azure Machine Learning NLP 最新動向 2020/07/02
DataDrift in Azure Machine Learning
Azure ML - OpenVINO 師走の推論ナイト 20191211 講演資料
AutoML & InterpretML (2019/11/27 Deep Learning Lab 講演資料)
品質管理 & 設備保全のための機械学習入門 2019年9月 (Quality Control and Predictive Maintenance App...
Neural Network Intelligence 概要 (AutoML Platform)
[DLLAB Engineer Days 講演資料] Automated ML & Model Interpretability

Recently uploaded (8)

PDF
世界半導体用酸化ハフニウム市場のサプライチェーン解析:上流、下流、収益モデル分析2025-2031
PDF
【QYResearch】グローバルコネクタ市場の動向と将来展望に関する詳細な分析報告
PDF
グローバルロープウェイ用スチールワイヤーロープ市場2025:主要企業のシェア、売上動向、競争戦略
PDF
【QYResearch】急拡大する医療機器市場における主要企業の動向と競争環境分析
PDF
XCMSを用いた質量分析データ処理_BioCAsia2021_yamamoto.pdf
PDF
【QYResearch】グローバル磁性材料産業チェーンの構造分析と市場動向の詳細解説
PDF
【QYResearch】グローバル農業機械市場の動向分析と成長戦略に関する総合調査報告
PDF
【QYResearch】人形ロボット産業の市場構造と今後の発展方向に関する分析レポート
世界半導体用酸化ハフニウム市場のサプライチェーン解析:上流、下流、収益モデル分析2025-2031
【QYResearch】グローバルコネクタ市場の動向と将来展望に関する詳細な分析報告
グローバルロープウェイ用スチールワイヤーロープ市場2025:主要企業のシェア、売上動向、競争戦略
【QYResearch】急拡大する医療機器市場における主要企業の動向と競争環境分析
XCMSを用いた質量分析データ処理_BioCAsia2021_yamamoto.pdf
【QYResearch】グローバル磁性材料産業チェーンの構造分析と市場動向の詳細解説
【QYResearch】グローバル農業機械市場の動向分析と成長戦略に関する総合調査報告
【QYResearch】人形ロボット産業の市場構造と今後の発展方向に関する分析レポート

Azure Machine Learning Build 2020