AI-901
Deep Dive
Azure Machine Learning is Microsoft's end-to-end platform for building, training, deploying, and managing ML models. The AI-901 tests its main components, when to use it, and how it differs from Foundry Tools (formerly Azure AI Services).
Contents
Azure ML is a cloud platform for the complete machine learning lifecycle: from exploring data to deploying models in production and monitoring them. It supports Python, R, and the most popular frameworks (TensorFlow, PyTorch, scikit-learn).
When to use Azure ML
The Workspace is the root resource of Azure ML — a workspace that holds all of the project's assets.
Workspace
The main container. Everything else (data, experiments, models, endpoints) lives inside a workspace.
Datastores
Connections to data storage (Azure Blob, Azure Data Lake, SQL). Training data is read from here.
Datasets / Data Assets
Registered versions of datasets. Enables reproducibility — knowing exactly which data each model was trained with.
Experiments / Jobs
Training runs. Each job logs metrics, parameters, logs, and the resulting model for comparison.
Model Registry
A catalog of trained models with versioning. Allows comparing versions and promoting the best model to production.
Environments
Definitions of the execution environment (Python dependencies, framework versions). Ensures reproducibility.
Create the project container where your datasets, experiments, models, and compute will live.
+1 more...
Compute Instance
An individual VM for development and interactive work. Like a Jupyter Notebook in the cloud with access to the workspace. For data exploration and code development.
Your laptop in the cloud — always on, always connected to the workspace.
Compute Cluster
A cluster of VMs that scales automatically from 0 to N nodes based on demand. For training models that require intensive or distributed compute.
Many laptops that appear when you need them and disappear when you're done. You only pay while training.
Serverless Compute
Azure fully manages the compute — you don't have to configure anything. Scales automatically. The simplest option for AutoML and Designer.
You don't know where it runs or how many machines are used — it just works.
Attached Compute
Connecting external compute (Azure Databricks, HDInsight, AKS, virtual machines) to the workspace. For special cases with existing infrastructure.
Bring your own hardware into the Azure ML workspace.
Create a GPU compute cluster that spins up only when there's a training job and spins back down on its own afterward.
+1 more...
Run a Python cell with the Azure ML SDK and compare this path against AutoML and Designer.
+1 more...
AutoML automates algorithm selection and hyperparameter tuning. You provide the data and define the objective (classification, regression, forecasting) — AutoML tries dozens of combinations and delivers the best model.
What does it automate?
Selecting the optimal algorithm (from a list of candidates)
Feature engineering (automatic feature engineering)
Data normalization and transformation
Hyperparameter tuning (search over the parameter space)
Model ensembling — combines the best models for higher accuracy
Supported task types
Classification
Predicting a category. E.g.: customer churn detection.
Regression
Predicting a number. E.g.: product price.
Time Series Forecasting
Predicting future values in time series. E.g.: inventory demand.
Computer vision
Image classification, object detection. With labeled data.
NLP
Text classification, summarization, information extraction.
Exam trap
AutoML doesn't eliminate the need for data scientists — it automates the algorithm-selection part, but you still need to prepare the data, define the success metric, interpret the results, and iterate. It's a productivity tool, not a complete replacement.
Let Azure automatically try 24 different algorithms and compare which one best predicts customer churn.
+1 more...
Designer is a drag-and-drop interface for building ML pipelines without writing code. You drag components (load data, transform, train, evaluate) and connect them visually.
Ideal for
Available components
Connect drag-and-drop modules — load data, clean, split, train, and evaluate — without writing a single line of code.
+1 more...
Once a model is trained, it's deployed as a REST endpoint so applications can consume it with real-time or batch predictions.
🚀 Online Endpoint
Predictions in real time. The client sends data → receives a prediction in milliseconds.
Use cases
📦 Batch Endpoint
Predictions in batches. Processes large volumes of data asynchronously — results are saved to storage.
Use cases
Turn your trained model into a REST API and test it live with a real request.
+1 more...
Tag and version a trained model so you can track it and roll back if a new version fails.
+1 more...
The most common AI-901 question
When should you use Azure ML and when should you use Foundry Tools?
| Criterion | Azure ML | Foundry Tools |
|---|---|---|
| Model type | Custom — you train with YOUR OWN data | Pre-trained — ready to use |
| Requires a data scientist | Yes (AutoML makes it easier but doesn't eliminate it) | No — just an API key |
| Training data | Yours — you need a labeled dataset | Microsoft's (trained at global scale) |
| Implementation time | Weeks/months (prepare data, train, evaluate) | Minutes (just call the API) |
| Typical use case | Churn prediction specific to your company | Translating text, detecting objects in photos |
| Cost | Per hour of training compute | Per number of API calls |
Quick rule: If the scenario says "your own data", "custom model", or "business-specific characteristics" → Azure ML. If it says "use AI for [general task]" without mentioning your own data → Foundry Tools.
Did you get this topic?
Test what you just learned
A startup wants to add automatic Spanish-to-English translation to its e-commerce app. It has no training data of its own and no data science team. Which Azure service is most appropriate?
Put what you learned to the test with original AI-901 exam questions.
Start AI-901 mock examUnlimited mock exams with the Pro plan
Log in to track your progress.