1. Azure Machine Learning Overview
1.1 Agenda
Estimated reading time: ~10 minutes
Learning Outcomes
- Define Azure Machine Learning and its role within the Azure AI ecosystem
- Describe the key components of the Azure ML workspace
- Distinguish (phân biệt) Azure ML from Azure AI prebuilt services
- Identify when to use Azure ML over prebuilt AI services
1.2 Glossary
| Term | Quick Explanation |
|---|---|
| Azure Machine Learning (Azure ML) | Dịch vụ đám mây được quản lý (managed cloud service) của Microsoft để xây dựng, huấn luyện (train), triển khai (deploy) và giám sát (monitor) các model ML tùy chỉnh. |
| Workspace (Không gian làm việc) | Đơn vị tổ chức (organizational unit) cấp cao nhất trong Azure ML — chứa mọi thứ: dữ liệu, thí nghiệm (experiments), model, và endpoints. |
| Compute (Tài nguyên tính toán) | Hạ tầng tính toán (computing infrastructure) — CPU/GPU cluster — dùng để chạy training và inference. |
| Pipeline | Luồng công việc tự động (automated workflow) kết nối các bước ML (data prep → train → evaluate) thành một chuỗi tái sử dụng (reusable). |
| MLOps | Machine Learning Operations — thực hành (practices) tích hợp DevOps vào ML để tự động hóa (automate), giám sát và quản lý vòng đời model trong production. |
| Model Registry (Kho model) | Nơi lưu trữ và quản lý phiên bản (version) các model đã được huấn luyện, kèm metadata và metrics. |
| Inference (Suy diễn) | Quá trình model đã được triển khai (deployed) áp dụng kiến thức đã học để đưa ra dự đoán trên dữ liệu mới. |
2. Problem Statement
Building custom ML models from scratch presents (đặt ra) several enterprise-level (cấp doanh nghiệp) challenges:
- Infrastructure complexity — Provisioning (cấp phát) GPUs, managing distributed training (huấn luyện phân tán), and scaling clusters require deep infrastructure expertise.
- Reproducibility (Khả năng tái hiện) — Without tracking, experiments are hard to reproduce (tái tạo) or audit (kiểm toán).
- Collaboration — Data scientists, ML engineers, and ops teams work with different tools, leading to silos (sự cô lập).
- Governance (Quản trị) — Production ML requires versioning, monitoring, and compliance (tuân thủ) — all operationally expensive (tốn kém) to build manually.
Azure Machine Learning addresses all four by providing a unified (thống nhất), managed platform for the entire ML lifecycle.
3. What is Azure Machine Learning?
3.1 Definition
Azure Machine Learning is a managed cloud service that provides the infrastructure (hạ tầng), tools, and workflows for teams to build, train, evaluate, deploy, and monitor custom Machine Learning models at scale (ở quy mô lớn).
3.2 Position in the Azure AI Ecosystem
3.3 Azure ML Workspace — Key Components
The workspace (không gian làm việc) is the root resource (tài nguyên gốc) that contains everything:
| Component | Description |
|---|---|
| Data assets (Tài sản dữ liệu) | Registered datasets with versioning and lineage (dõi dấu vết) tracking |
| Compute targets | CPU/GPU clusters for training; inference clusters for deployment |
| Experiments & runs (Thí nghiệm và lần chạy) | Tracked training runs with metrics, logs, and artifacts (kết quả) |
| Model registry | Versioned (có phiên bản) catalog of trained models with metadata |
| Endpoints | REST API endpoints (điểm cuối API) for real-time or batch inference |
| Pipelines | Reusable, automated (tự động hóa) ML workflow definitions |
| Environments (Môi trường) | Containerized (đóng gói) dependency configurations for reproducibility |
4. Azure ML vs. Prebuilt AI Services
This is the most critical distinction (sự phân biệt quan trọng nhất) for AI-900:
| Dimension | Prebuilt AI Services | Azure Machine Learning |
|---|---|---|
| What you get | Ready-to-call API for common tasks | Platform to build your own models |
| Training required | No — model already trained (đã huấn luyện sẵn) | Yes — you train on your own data |
| Customization | Limited (fine-tuning only for some services) | Full control over algorithm, features, architecture |
| Data needed | None (for standard features) | Your own labeled dataset |
| Expertise required | Minimal ML knowledge | Data science skills needed |
| Time to implement | Hours to days | Weeks to months |
| Best for | Common tasks: sentiment, OCR, translation | Domain-specific (chuyên biệt) problems not covered by prebuilt services |
4.1 Decision Rule
5. Benefits of Azure Machine Learning
| Benefit | Detail |
|---|---|
| Scalability (Mở rộng) | Automatically provision (cấp phát) and release GPU clusters — no idle (nhàn rỗi) infrastructure costs |
| Experiment tracking | Log metrics, parameters, and model artifacts automatically across every training run |
| Collaboration | Shared workspace for teams — data scientists, ML engineers, and stakeholders (các bên liên quan) |
| Responsible AI | Built-in Fairness, Explainability (khả năng giải thích), and Error Analysis dashboards |
| Security & Compliance | RBAC (Role-Based Access Control), private networking (mạng riêng tư), audit logs, GDPR/HIPAA support |
| MLOps | CI/CD (Continuous Integration/Deployment) pipelines for automated model retraining and promotion (thăng hạng model) |
6. Discussion Questions
Q1 — The Right Tool: A logistics company wants to predict package delivery times using 5 years of historical shipping data (origin, destination, package weight, carrier, weather). The data has 12 million records with a custom feature set unique to their network. Should they use Azure AI Language, Azure ML, or Azure OpenAI Service? Justify with specific (cụ thể) reasons.
Q2 — MLOps Necessity: A bank deploys a credit scoring model using Azure ML. The model was accurate when deployed in January. By August, the model's predictions diverge (sai lệch) noticeably from actual outcomes, but no one detects (phát hiện) this until customers complain (phàn nàn). What MLOps practices should have been in place? What specific Azure ML features address each practice?
Q3 — Cost vs. Control: A startup is choosing between Azure AI Language (prebuilt sentiment API at $1/1000 calls) and Azure ML (custom sentiment model requiring 2 months of development + GPU training costs). The startup processes 500,000 reviews per month, but their reviews are highly domain-specific (chuyên biệt) — technical jargon (thuật ngữ chuyên ngành) the prebuilt model misclassifies 25% of the time. How would you frame (đặt vấn đề) the build-vs-buy analysis?
Made by Anh Tu - Share to be share