2.6 Azure Machine Learning Service
Key Takeaways
- Azure Machine Learning is a cloud platform for training, deploying, and managing ML models — it provides tools for the entire ML lifecycle.
- Automated ML (AutoML) automatically selects the best algorithm and hyperparameters for your data — no ML expertise required.
- Azure ML Designer provides a drag-and-drop visual interface for building ML pipelines without writing code.
- Azure ML supports model deployment as real-time endpoints (for immediate predictions) and batch endpoints (for processing large datasets).
- The AI-900 tests your awareness of Azure ML capabilities, not your ability to use the platform or write code.
Quick Answer: Azure Machine Learning is a cloud platform for the entire ML lifecycle. Key features include Automated ML (AutoML) for automatic algorithm selection, Designer for visual drag-and-drop pipeline building, and model deployment as real-time or batch endpoints. Azure ML workspace is the central resource that organizes data, compute, models, and experiments.
What Is Azure Machine Learning?
Azure Machine Learning is Microsoft's cloud-based platform for building, training, deploying, and managing machine learning models. It provides tools for every stage of the ML lifecycle:
| Lifecycle Stage | Azure ML Feature | Description |
|---|---|---|
| Data preparation | Data assets, Datastores | Connect to data sources, clean and transform data |
| Model training | AutoML, Designer, Notebooks | Train models using automated, visual, or code-based approaches |
| Model evaluation | Metrics, Responsible AI dashboard | Evaluate model performance and fairness |
| Model deployment | Endpoints | Deploy models as real-time or batch endpoints |
| Model monitoring | Model monitoring | Track model performance and detect drift in production |
Azure ML Workspace
The Azure Machine Learning workspace is the top-level resource that organizes all ML assets:
| Asset | Description |
|---|---|
| Datastores | Connections to data storage (Azure Blob, Data Lake, SQL) |
| Data assets | Registered datasets used for training and evaluation |
| Compute | Computing resources for training (clusters, instances) |
| Environments | Software configurations (Python packages, Docker images) |
| Models | Trained ML models registered for deployment |
| Endpoints | Deployed model endpoints for serving predictions |
| Pipelines | Automated workflows for data processing and model training |
| Experiments | Collections of training runs with metrics and logs |
Automated ML (AutoML)
Automated Machine Learning (AutoML) is one of the most important Azure ML features for the AI-900 exam. It automates the process of selecting the best ML algorithm and hyperparameters for your data.
What AutoML Does
- Receives your dataset — you provide training data with features and labels
- Tries multiple algorithms — linear regression, decision trees, random forests, gradient boosting, neural networks, etc.
- Tunes hyperparameters — automatically adjusts settings for each algorithm
- Evaluates performance — compares all models using appropriate metrics
- Recommends the best model — presents the top-performing model with its metrics
AutoML Task Types
| Task Type | Output | Example |
|---|---|---|
| Classification | Category | Predict spam/not spam |
| Regression | Number | Predict house price |
| Time-series forecasting | Future numbers | Predict next month's sales |
Benefits of AutoML
- No ML expertise needed — the system selects the algorithm for you
- Fast experimentation — test dozens of algorithms in minutes
- Transparent results — see which algorithms were tried and their performance
- Reproducible — all experiments are logged and can be replicated
On the Exam: AutoML is a frequent topic. Know that it automatically selects the best algorithm and hyperparameters, requires no coding or ML expertise, and supports classification, regression, and time-series forecasting tasks.
Azure ML Designer
Azure ML Designer provides a drag-and-drop visual interface for building ML pipelines without writing code.
How Designer Works
- Drag dataset onto the canvas
- Add processing components — data transformations, feature selection
- Add training component — select an algorithm
- Connect components — create a visual pipeline flow
- Run the pipeline — execute training on Azure compute
- Evaluate results — view metrics and model performance
Designer vs. AutoML vs. Notebooks
| Approach | User Type | Code Required | Best For |
|---|---|---|---|
| AutoML | Business users, beginners | No | Quick model building, no ML expertise |
| Designer | Data analysts, visual learners | No | Custom pipelines with visual interface |
| Notebooks | Data scientists, developers | Yes (Python/R) | Full control, custom algorithms, experimentation |
Compute Resources
Azure ML provides different compute types for different needs:
| Compute Type | Purpose | When to Use |
|---|---|---|
| Compute instance | Development VM for notebooks | Interactive development and testing |
| Compute cluster | Scalable cluster for training | Large-scale model training (scales to zero when idle) |
| Serverless compute | On-demand training compute | Quick training jobs without managing infrastructure |
| Kubernetes | Inference cluster | Real-time model serving in production |
Model Deployment
After training, you deploy models as endpoints for applications to consume:
Real-Time Endpoints
- Purpose: Immediate predictions (milliseconds response time)
- Use case: A web app that predicts house prices when a user enters property details
- How it works: Send a request with features, get a prediction back instantly
Batch Endpoints
- Purpose: Process large datasets (minutes to hours)
- Use case: Score 1 million customer records for churn prediction overnight
- How it works: Submit a dataset, results are stored when processing completes
On the Exam: Know the difference between real-time (immediate predictions) and batch (large dataset processing) endpoints. A question asking "How should you serve predictions to a mobile app?" → real-time endpoint. "How should you score 500,000 records?" → batch endpoint.
What does Azure Automated ML (AutoML) automatically do?
Which Azure ML feature provides a drag-and-drop visual interface for building ML pipelines without code?
A mobile app needs to get house price predictions immediately when a user enters property details. Which deployment type should you use?
Which THREE task types does Azure Automated ML (AutoML) support? (Select three)
Select all that apply
MLOps: Managing Models in Production
Azure Machine Learning is not just for training — it supports MLOps, the practice of operating models reliably after deployment. The AI-900 skills measured explicitly call out "model management and deployment capabilities," so know these building blocks:
| Capability | What It Does |
|---|---|
| Model registry | Versions trained models so you can track, compare, and roll back |
| Environments | Captures the exact Python packages and runtime so results are reproducible |
| Pipelines | Chain data prep, training, and scoring into a repeatable, automated workflow |
| Model monitoring | Watches a live model for data drift — when production data drifts away from the training data and accuracy degrades |
| Endpoints | Serve predictions in real time or in batch |
On the Exam: "Data drift" is the reason a deployed model's accuracy decays over time, and the remedy is to retrain on fresh data. Model monitoring is the capability that detects it.
Compute, Quickly Distinguished
Learners mix up the compute types. The clean split is development vs. training vs. serving:
- Compute instance — a single managed VM for interactive notebook development by one data scientist.
- Compute cluster — a multi-node, auto-scaling cluster for training jobs; it can scale to zero when idle to save money.
- Inference endpoint (managed online or Kubernetes) — for serving a deployed model's predictions.
Three Authoring Experiences, One Decision
Azure ML offers three ways to build a model. Pick by the user's skills and need for control:
| If the user... | Recommend |
|---|---|
| Has no ML or coding skills and wants the best model fast | Automated ML (AutoML) |
| Wants a no-code visual pipeline they can customize | Designer |
| Is a data scientist who needs full control and custom code | Notebooks (Python/R) |
Endpoint Selection Cheat Sheet
| Requirement | Endpoint Type |
|---|---|
| Instant, single-request predictions for an app or website | Real-time (online) endpoint |
| Score a large file or table on a schedule | Batch endpoint |
| Low-latency responses to many concurrent users | Real-time with autoscaling |
| Overnight scoring of millions of records, cost-sensitive | Batch |
On the Exam: The classic pairing — interactive app needs predictions now → real-time endpoint; score a big dataset where latency does not matter → batch endpoint. And remember the value proposition of AutoML: it removes the need for ML expertise by automatically selecting the algorithm and tuning hyperparameters for classification, regression, and time-series forecasting.