AI-900 Microsoft Azure AI Fundamentals – Key Concepts

These question-and-answer flashcards cover major concepts, services, responsible-AI practices, machine-learning techniques, and common workloads required for the Microsoft Azure AI Fundamentals (AI-900) exam.

What are the three most common Azure AI workload categories covered in AI-900?

Computer vision, natural language processing (including speech), and knowledge mining.

Which Azure service provides pre-built APIs for vision, speech, language, and decision tasks?

Azure AI Services (formerly Cognitive Services).

Name two core capabilities offered by Azure AI Vision.

(1) Image analysis with tagging, captioning, smart-cropping, and OCR; (2) Face detection / recognition and spatial analysis on video streams.

What service helps filter text, images, and video for harmful or unsafe content?

Azure AI Content Safety (successor to Azure Content Moderator).

Which Azure service automates text, key-value pair, and table extraction from documents?

Azure AI Document Intelligence (formerly Form Recognizer).

Define knowledge mining in Azure.

Using services like Azure AI Search plus AI skills to ingest, enrich, and explore large volumes of heterogeneous content for hidden insights.

List Microsoft’s six Responsible AI principles.

Fairness, reliability & safety, privacy & security, inclusiveness, transparency, and accountability.

What two primary harms does machine-learning fairness aim to avoid?

Harm of allocation and harm of quality-of-service.

Which open-source toolkit is integrated with Azure ML for fairness assessment and mitigation?

Fairlearn.

In Azure’s AI shared-responsibility model, who is responsible for protecting against harmful prompts sent to a PaaS model?

Microsoft (the platform provider) implements input & output safety systems, while users remain responsible for application-level safeguards.

What is the purpose of Microsoft’s Transparency Notes?

To explain how an AI system works, its capabilities, limitations, and risks so customers can make informed deployment decisions.

Give three tasks that AutoML in Azure Machine Learning can automate.

Algorithm selection, hyper-parameter tuning, and automated feature pre-processing.

Which compute resource scales up training jobs automatically in Azure ML?

A Compute Cluster.

Deep learning differs from traditional ML primarily because it …

uses multi-layer neural networks that learn hierarchical representations directly from unstructured data.

What are labeled datasets in Azure ML used for?

Supervised training, validation, and testing; they contain input data plus ground-truth labels created via data-labeling projects.

Explain K-fold cross-validation in one sentence.

The data is split into K subsets; K models are trained each using K-1 folds for training and 1 fold for validation, and results are averaged.

Which two solution experiences can you choose when starting an AutoML run?

Code-first (SDK/CLI) or no-code (Azure ML Studio UI).

What query parameter must be added to Azure AI Vision REST calls to enable object detection?

visualFeatures=Objects

State one limitation of Azure AI Vision object detection.

Objects smaller than ~5 % of the image may not be detected reliably.

Which API detects human faces and returns attributes like age, emotion, and landmarks?

Azure AI Face service (Detect API).

What is the main difference between image classification and object detection?

Classification labels the entire image; object detection finds and labels individual objects with bounding boxes.

Name two language-centric pre-configured features in Azure AI Language.

Key-phrase extraction and sentiment analysis / opinion mining.

What are the two modes of Azure AI Translator Document Translation?

Asynchronous batch translation and synchronous single-document translation.

Which deployment options does Azure AI Speech support besides the public cloud?

Edge/on-prem containers and sovereign clouds.

Give three typical use-cases for speech-to-text offered by Azure.

Real-time call-center transcription, captioning meetings, voice-assistant commands (many more acceptable).

What is Generative AI?

A branch of AI that creates new content (text, images, code, etc.) from learned patterns, rather than just analysing existing data.

List four common use-cases for generative AI on Azure OpenAI Service highlighted by Microsoft.

Content creation, chatbots/virtual assistants, code generation, and personalized marketing (among others).

Transformer-based language models learn by predicting .

the next token (word/piece) in large text corpora, capturing semantic relationships.

Compare LLMs and SLMs in one key aspect.

LLMs have billions of parameters and broad knowledge; SLMs are smaller, topic-focused, and easier to fine-tune locally.

What tool in Azure ML streamlines prompt engineering and evaluation workflows?

Prompt Flow.

Which integrated service scores generated content for hate, sexual, violent, or self-harm risk?

Azure AI Content Safety (used by Azure AI Studio safety evaluations).

In Azure OpenAI, what must you create before issuing inference calls?

A deployment of the chosen model within the Azure OpenAI resource.

Give one security advantage Azure OpenAI has over public OpenAI APIs.

Private networking and Azure’s enterprise-grade compliance/security controls (e.g., VNet integration, regional isolation).

What metric does linear regression try to minimize when using Ordinary Least Squares?

The sum of squared differences between observed and predicted values (residuals).

Name three supervised learning algorithms besides linear regression.

Logistic regression, decision trees, support vector machines (others acceptable: random forest, naïve Bayes, KNN).

Which Azure service provides semantic search, indexing, and AI-powered data enrichment?

Azure AI Search (formerly Cognitive Search).

What is MLOps and why is it needed?

A set of practices for continuous integration, delivery, monitoring, and governance of ML models to ensure reliable, repeatable production use.

List the five lifecycle stages in MLOps.

Development, testing, deployment, monitoring, and maintenance.

How does Azure ML assist with model interpretability?

It provides Responsible AI dashboard tools for feature importance, error analysis, fairness, and explanation of predictions.

What does ARIMA stand for in time-series forecasting?

AutoRegressive Integrated Moving Average.

Give two ensemble learning techniques.

Random forest and boosting methods such as AdaBoost or XGBoost.

What Azure service combines ingestion, storage, and big-data analytics and often hosts Spark NLP workloads?

Azure Databricks or Azure Synapse (either earns credit).

Which NLP task turns spoken audio into written text?

Speech recognition (speech-to-text).

What REST feature would you call in Azure AI Language to detect a document’s language?

Language Detection.

Which Azure AI Language feature links entities to Wikipedia URLs?

Entity Linking.

What is a custom NER model used for?

Recognising domain-specific entities not covered by the pre-built NER list.

Describe the purpose of a data-split between training, validation, and test sets.

To train the model, tune hyper-parameters without overfitting, and finally obtain an unbiased estimate of performance.

Why is stratified K-fold cross-validation preferable on imbalanced datasets?

It preserves the class distribution in each fold, giving more reliable estimates.

State one benefit and one drawback of large language models.

Benefit: Versatile, high-quality language generation; Drawback: High compute cost and potential latency/local-deploy challenges.

Which Microsoft principle focuses on explaining model decisions to users?

Transparency.

How does liveness detection improve face verification security?

By ensuring the subject is a live person rather than a photo, video, or mask presented to the camera.

What parameter lets you request face attributes like age or emotion in the Face Detect API?

returnFaceAttributes.

Name three computer-vision use-cases listed for OCR.

Digitizing documents, reading street signs/product labels, extracting form data (any similar valid example).

What Azure Vision feature automatically writes captions for pictures?

Image analysis with caption generation (part of Azure AI Vision).

Which document types can Azure Document Translation handle while preserving layout?

PDF, Word (doc/docx), PowerPoint (ppt/pptx), Excel (xls/xlsx), CSV, and more—maintaining original structure.

Explain the harm of allocation with an example.

An AI system unfairly denies loans to a protected group, allocating financial resources unequally.

What is the purpose of ‘groundedness evaluation’ in Azure AI Studio safety tests?

To detect unsupported (hallucinated) portions of generated answers so they can be corrected.

Which Azure ML component allows drag-and-drop pipeline authoring?

Azure ML Designer.

Why might you choose gradient-descent training over OLS in linear regression?

Gradient descent scales better to very large datasets or when data can’t fit in memory.

What is a Compute Instance in Azure ML (contrast with Compute Cluster)?

A single, dedicated VM for interactive development; clusters autoscale multiple nodes for batch jobs.

Which service would you use to build a bot interface that queries a knowledge-mined index?

Azure AI Bot Service connected to Azure AI Search index.

Name two deployment targets supported by Azure ML for real-time inference.

Azure Kubernetes Service (AKS) and Managed Online Endpoints (or Azure Functions for small workloads).

What is prompt engineering?

Crafting and refining the input text prompts sent to generative models to elicit desired, high-quality outputs.

Which Azure feature enables on-prem, air-gapped deployment of vision or speech models?

Containerized Cognitive Services (distroless Docker containers).

Identify one challenge unique to computer-vision workloads compared with NLP.

High variability in lighting, angle, occlusion, or scale requiring robust image pre-processing and larger data volumes.

What capability does Spatial Analysis in Azure AI Vision provide?

Real-time people counting and movement tracking in video streams for scenarios such as retail analytics or safety monitoring.

How can Azure AI Face ‘Find Similar’ be used?

Given a target face, it returns similar faces from a supplied set—useful for photo deduplication or face search.

Give one reason to choose synchronous document translation over asynchronous.

Immediate translation of a single small file without needing Azure Blob Storage.

What is a semantic search advantage over keyword search?

It understands user intent and context to return conceptually relevant results, not just exact keyword matches.

Describe ‘intelligent stopping’ in Azure AutoML.

The system ends an experiment early when further iterations are unlikely to improve the metric, saving compute cost.

Which metrics would you inspect for fairness disparity in a binary classifier?

Accuracy, precision, recall, F1, false-positive/negative rates across sensitive subgroups.

What is ONNX Runtime used for in Azure?

Accelerating training and inference of deep-learning models across multiple hardware targets.

Name two industries highlighted for AI-powered image classification in Azure.

Retail fashion and insurance/finance (claims assessment).

What aspect of an AI workload is the customer always responsible for regardless of SaaS/PaaS/IaaS?

Using the system ethically and evaluating outputs within their business context (final human oversight).

Which feature of Azure AI Speech helps build a unique brand voice?

Custom Neural Voice (text-to-speech voice cloning).

Explain the role of ‘precision’ in an object detector.

Precision measures the proportion of detected objects that are true positives (correct).

What is differential privacy and where is it applied in Azure ML?

A technique to add statistical noise and prevent re-identification of individuals in datasets, supported in responsible data handling.

List three built-in governance capabilities Azure AI offers for Responsible AI.

Content filtering, safety evaluations, audit trails/datasheets, model monitoring dashboards (any three).

Why might an organisation deploy generative AI models on Azure Arc?

To run and manage the models consistently across on-prem, multi-cloud, or edge environments with unified governance.

Which Azure OpenAI model converts speech to text?

Whisper (speech-to-text model).

What are the two main approaches to handle multilingual NLP in Azure?

Use pre-built multilingual models or train once in one language with transfer to others.

What capability does ‘speaker recognition’ add on top of speech-to-text?

Identifies or verifies who is speaking, not just what is being said.

Define ‘ensemble learning’ in one sentence.

Combining predictions from multiple base models to improve accuracy and robustness over any single model.

How does Boosting differ from Bagging?

Boosting trains models sequentially focusing on previous errors; bagging trains models in parallel on random subsets (e.g., random forest).

Which metric indicates over-prediction of a minority class in classification?

False-positive rate (or precision drop) for that class.

Provide an example of contextual understanding in generative AI.

A chatbot remembers earlier parts of the conversation to answer follow-up questions coherently.

Why is GPU acceleration important for deep learning on Azure?

Parallel processing of matrix operations drastically reduces training time for large neural networks.

What is a ‘messyGroup’ in Azure Face Group API output?

An array containing faces that didn’t fit into any similarity group—i.e., no clear match found.

Which algorithm family is commonly used for unsupervised dimensionality reduction?

Principal Component Analysis (PCA).

What does ‘ground-truth’ mean in labeled datasets?

The correct, authoritative label or value against which model predictions are compared.

Why is data drift monitoring critical after model deployment?

To detect when incoming data distribution changes, potentially degrading model accuracy.

Describe the purpose of Azure AI Vision’s ‘custom image classification’ in one line.

Let users train specialized classifiers with their own labeled images to recognise domain-specific categories.

In Microsoft’s Responsible AI Lifecycle, what activity follows risk identification?

Measurement/quantification of harms, then mitigation strategies and ongoing monitoring.

Which feature of Azure Bot Service can integrate generative models for rich responses?

DirectLine or Azure OpenAI connector extensions enabling GPT-powered conversations.

How can developers test Azure AI Vision quickly without writing code?

By using Vision Studio web interface to upload images or video URLs and view results.

What is the primary purpose of Azure Machine Learning?

To provide a cloud-based platform for building, training, deploying, and managing machine learning models at scale.

Which Azure AI Language feature or service can be used for text-based conversational interfaces like chatbots by understanding user intent or answering questions?

Azure AI Language's Conversational Language Understanding (CLU) or Question Answering (QnA), often integrated with Azure AI Bot Service.

Name an Azure AI service focused on detecting statistical abnormalities in data streams.

Azure AI Anomaly Detector.

What is the main goal of Azure AI

To deliver personalized, relevant content and experiences to users in real-time by learning from user behavior and content.