Machine Learning Fundamentals

Why Machine Learning Matters for AI-901

Module 1 introduced AI as software that learns from data. Every workload in this guide (generative AI, NLP, speech, computer vision, information extraction) is built on machine learning under the hood. The AI-901 exam tests whether you understand that foundation, not just the surface-level capabilities.

This module covers the ML concepts the exam returns to repeatedly: the three types of learning, the specific tasks within each, how models are trained and evaluated, and what Azure provides for building ML solutions. Read this before the workload-specific modules and the references to "training data", "classification", and "model evaluation" throughout the guide will make immediate sense. The concepts here also underpin the responsible AI tools you will encounter in Module 8, particularly the Responsible AI dashboard.

How to use this module: The AI-901 exam directly tests the three types of ML and their tasks (supervised, unsupervised, reinforcement), Azure Machine Learning capabilities (AutoML, Designer, managed endpoints), and the Azure ML vs. Azure AI services distinction. Sections on training mechanics, evaluation metrics, and overfitting are useful context that helps you reason through exam scenarios - but they are tested at a conceptual level only, not at a mathematical or implementation depth. If you are short on time, prioritise the three types of ML, Azure ML capabilities, and the final Key Takeaways.

---

The Three Types of Machine Learning

Machine learning comes in three broad forms, each suited to a different kind of problem.

Supervised Learning

In supervised learning, the model trains on labeled examples: input data paired with the correct output. A dataset of house photos labeled "cat" or "no cat", or a set of house listings each labeled with its sale price, are both supervised learning datasets.

The model learns to map inputs to outputs by seeing thousands of these examples. At inference time, it applies that learned mapping to new inputs it has never seen before.

Supervised learning powers most of the AI workloads in this guide: image classification, sentiment analysis, speech recognition, and named entity recognition are all supervised learning tasks.

Supervised learning breaks into two major task types:

Classification predicts which category an input belongs to. The output is a discrete label. Examples:

• Is this email spam or not spam? (binary classification, two possible labels)
• Which digit is in this image: 0 through 9? (multiclass classification, ten possible labels)
• Which topics apply to this news article? (multilabel classification, multiple labels can apply)

Regression predicts a continuous numerical value. The output is a number, not a category. Examples:

• What will this house sell for?
• How many units will we sell next month?
• What is the predicted temperature tomorrow?

The exam distinguishes these clearly. If the output is a category, it is classification. If the output is a number on a continuous scale, it is regression.

Unsupervised Learning

In unsupervised learning, the model trains on data with no labels. There are no correct answers to learn from. Instead, the model discovers patterns and structure in the data on its own.

The most important unsupervised task for the exam is clustering: grouping similar items together based on their features, without being told in advance what the groups should be. Examples:

• Grouping customers into segments based on purchasing behavior
• Grouping documents by topic without pre-defined topic categories
• Identifying unusual patterns in network traffic (anomaly detection)

A related technique is dimensionality reduction, which simplifies data by finding the most important underlying patterns. This is used to compress data before training or to visualize high-dimensional data in two or three dimensions.

The key distinction from supervised learning: unsupervised models do not predict a known output. They find structure in data where no labels exist.

Reinforcement Learning

In reinforcement learning, a model (called an agent) learns by interacting with an environment. It takes actions, receives rewards for good outcomes and penalties for bad ones, and gradually learns a policy that maximizes cumulative reward.

Reinforcement learning is behind game-playing AI (AlphaGo, chess engines), robotics control, and some components of how large language models are fine-tuned (a technique called RLHF: Reinforcement Learning from Human Feedback). For the AI-901 exam, you need to know what it is and one or two examples, but not the mathematical details.

---

How Models Are Trained

Understanding the training process helps you answer exam questions about data requirements, overfitting, and model evaluation.

Features and Labels

Features are the input variables the model uses to make predictions. For a house price prediction model, features might be: square footage, number of bedrooms, location (as a numeric code), and year built.

Labels are the correct output values in a supervised dataset: the actual sale price for each house, or "cat"/"not cat" for each image.

Feature engineering is the process of selecting, transforming, and creating features from raw data to make them more useful to the model. For example, converting a raw date into day-of-week and month features.

Training, Validation, and Test Splits

A labeled dataset is split into three parts before training:

Split	Purpose	Typical Size
Training set	The data the model learns from	70-80%
Validation set	Used during training to tune the model and avoid overfitting	10-15%
Test set	Held out until the very end to give an unbiased measure of final performance	10-15%

The test set must never be used during training or tuning. Using it early contaminates the final performance measurement.

Overfitting and Underfitting

Overfitting happens when a model learns the training data too well, including its noise and quirks, and fails to generalize to new data. The training accuracy is high but performance on the validation or test set is much lower. The model has memorized the training examples rather than learning the underlying pattern.

Underfitting happens when the model is too simple to capture the patterns in the data. Both training and test performance are poor.

The goal is a model that performs well on data it has never seen: good generalization.

Model Evaluation Metrics

Different tasks use different metrics:

For classification:

• Accuracy: the percentage of predictions that are correct. Simple but misleading when classes are imbalanced (if 99% of emails are not spam, a model that predicts "not spam" for everything has 99% accuracy but is useless).
• Precision: of all the items the model labeled as positive, how many actually were? High precision means few false positives.
• Recall: of all the actual positive items, how many did the model find? High recall means few false negatives.
• F1 score: the harmonic mean of precision and recall, useful when you need to balance both.
• AUC-ROC: measures how well the model separates classes across all classification thresholds.

For regression:

• Mean Absolute Error (MAE): average of absolute differences between predicted and actual values
• Root Mean Squared Error (RMSE): similar to MAE but penalizes larger errors more heavily
• R-squared (R2): proportion of variance in the target explained by the model (1.0 is perfect, 0 is no better than predicting the mean)

---

Azure Machine Learning

The Azure service specifically for machine learning is Azure Machine Learning (Azure ML). It is a managed cloud platform for building, training, deploying, and monitoring machine learning models.

Azure ML is distinct from the AI services covered in the workload modules (Azure AI Vision, Azure AI Speech, etc.). Those are pre-built AI services: you call them with an API and they return results. Azure ML is for when you want to build and train your own models from scratch, or fine-tune existing ones on your own data.

Key Azure ML Capabilities

Automated ML (AutoML) trains and evaluates many different model types and hyperparameter combinations automatically, then recommends the best-performing one. You provide labeled data and specify the task type (classification, regression, or time-series forecasting) and AutoML handles the rest.

Azure ML Designer is a visual drag-and-drop interface for building ML pipelines without writing code. You connect components (data ingestion, feature engineering, model training, evaluation) into a workflow. This is useful for the exam because it represents a no-code/low-code path to ML.

Compute resources in Azure ML handle the actual training. Compute clusters scale automatically based on job demand. Compute instances are individual VMs for development and experimentation.

Model registry stores versioned model artifacts. Once a model is trained and evaluated, it is registered and can then be deployed to an endpoint.

Endpoints in Azure ML are where deployed models receive inference requests. A real-time endpoint (online endpoint) handles individual requests as they arrive. A batch endpoint processes large datasets asynchronously.

The Responsible AI dashboard in Azure ML provides visual tools for analyzing model fairness, error analysis, and explainability. It helps you understand where a model makes mistakes and whether those mistakes are distributed unevenly across demographic groups, directly supporting the fairness and accountability principles. Module 8 revisits this dashboard in full depth, showing exactly which exam scenarios it addresses.

Azure ML vs Foundry: Which Is Which?

This distinction matters for the exam, and Module 3 goes deep on everything Foundry offers once you have these ML foundations in place:

	Azure Machine Learning	Microsoft Foundry
Primary use	Build and train your own models	Deploy and use existing models (including LLMs)
Who uses it	Data scientists building custom models	Developers building AI applications
Key tools	AutoML, Designer, compute clusters, Responsible AI dashboard	Model catalog, Agent Builder, Foundry IQ, Prompt Flow
Models	Your own trained models	Pre-built models from Microsoft, OpenAI, Anthropic, and others

Both live in Azure, both are accessed through their respective portals, and both support deployment. The key difference is whether you are building a model yourself or using one that already exists.

---

Key Takeaways for the Exam

• The three types of machine learning are: supervised (labeled data, learns to predict known outputs), unsupervised (unlabeled data, finds patterns), and reinforcement (learns by trial and error with rewards).
• Supervised learning tasks: classification (predicts a category) and regression (predicts a number).
• Unsupervised learning tasks: clustering (groups similar items) and dimensionality reduction.
• Features are input variables; labels are correct output values in supervised datasets.
• Datasets are split into training, validation, and test sets. The test set is held out until final evaluation.
• Overfitting: model performs well on training data but poorly on new data. Underfitting: model performs poorly on both.
• Classification metrics: accuracy, precision, recall, F1, AUC-ROC. Regression metrics: MAE, RMSE, R-squared.
• Azure Machine Learning is for building and training your own models (AutoML, Designer, compute clusters, Responsible AI dashboard). Microsoft Foundry is for deploying and using pre-built models.
• AutoML automates model selection and hyperparameter tuning. The Responsible AI dashboard analyzes fairness and model errors.