6. Only a small fraction of real-world ML systems is composed of the ML code, as shown by the small blue box in the
middle. The required surrounding infrastructure is vast and complex.
ML
Code
Configuration
Data Collection
Data
Verification
Feature
Extraction
Machine
Resource
Management
Analysis Tools
Process
Management Tools
Serving
Infrastructure
Monitoring
“Hidden Technical Debt in Machine Learning Systems,” Google NIPS 2015
32. Model reproducibility Model retrainingModel deploymentModel validation
Train model Validate
model
Deploy
model
Monitor
model
Build appCollaborate Test app Release app Monitor app
App developer
using Azure DevOps
Data scientist using
Azure Machine Learning
Retrain model
Azure Machine Learning extension
for Azure DevOps
Data
(Model)
Code
機械学習はコードのみならずデータやモデルを管理する仕組みが必要。再現可能な環境を整えて
おくことで、モデルのライフサイクルを継続的に回すことができる。
45. Frameworks Azure
Machine Learning Operations
Services
オンプレミス
Azure Machine Learning
Ubuntu VM
Windows Server 2019 VM
Azure Custom Vision Service
ONNX Model
アプリケーション(C#, C, Javascript)
エッジ & IoT デバイズ
ONNX Runtime is open source
ML.NET
Automated Machine Learning
49. Microsoft Azure
Azure StorageService Endpoint
AML Compute
AML Service
Customer VNet
Compute Instance
Azure Key Vault
Service Tags
AKS Cluster
On-premises
VPN Gateway
Express Route
ExpressRoute public peering or
Internet Access through NAT IPs
Customer VNet
55. Ingest Store Prep & train Model & serve
Azure Blob Storage
Logs
(unstructured)
Azure Data Factory
Microsoft Azure also supports other Big Data services like Azure HDInsight and Azure Data Lake
to allow customers to tailor the above architecture to meet their unique needs.
Media
(unstructured)
Files
(unstructured)
Polybase
Business/custom
apps
(structured)
Azure SQL
Data
Warehouse
Azure
Analysis
Services
Power BI
Azure ML
Azure DevOps
Azure
Databricks