8. Comprehensive Glossary
This glossary provides concise definitions for all technical terms used throughout the guide. Terms are organized alphabetically with cross-references to relevant sections and flags for frequently tested concepts.
Legend: ⚠️ = Frequently tested on exam | → = See section reference
A
Accountability — Responsible AI principle ensuring AI systems meet legal and ethical standards with clear human responsibility. → Section 2.2.1
Automated ML (AutoML) ⚠️ — Azure capability that automatically selects, trains, and evaluates machine learning algorithms. Eliminates manual algorithm selection but still requires data preparation. → Section 3.3.1
Azure AI Face ⚠️ — Azure service for facial detection and analysis. Detection and attributes (age, emotion) are generally available; recognition features require Limited Access approval. → Section 4.2.2
Azure AI Foundry ⚠️ — Microsoft's unified platform for building AI applications with access to multiple models, development tools, and governance features. Formerly known as Azure AI Studio. → Section 6.2.1
Azure AI Language ⚠️ — Azure service providing text analysis capabilities including sentiment analysis, entity recognition, key phrase extraction, and language detection. → Section 5.2.1
Azure AI Speech ⚠️ — Azure service for speech-to-text, text-to-speech, speech translation, and speaker recognition. Note: Speech translation is here, NOT in Azure Translator. → Section 5.2.2
Azure AI Vision ⚠️ — Azure service providing image analysis capabilities including OCR, object detection, image classification, and image tagging. → Section 4.2.1
Azure Machine Learning — Azure platform for building, training, and deploying custom machine learning models with AutoML, Designer, and SDK options. → Section 3.3
Azure OpenAI Service ⚠️ — Azure service providing enterprise access to OpenAI models (GPT-4, GPT-3.5, DALL-E, Whisper) with added security, compliance, and regional availability. → Section 6.2.2
Azure Translator ⚠️ — Azure service for TEXT-to-text translation only. Does NOT handle speech—use Azure AI Speech for spoken language translation. → Section 5.1.6
B
Binary Classification ⚠️ — Classification with exactly two possible outcomes (yes/no, true/false, spam/not spam). A subset of classification, not regression. → Section 3.1.2
Bounding Box ⚠️ — Rectangle coordinates identifying the location of an object in an image. Key differentiator: object detection returns bounding boxes; classification does not. → Section 4.1.2
C
Classification ⚠️ — Supervised learning technique that predicts categorical labels (discrete categories), not numbers. Includes binary and multiclass variants. → Section 3.1.2
Clustering ⚠️ — Unsupervised learning technique that groups similar data points without predefined labels. No labels = unsupervised = clustering. → Section 3.1.3
Computer Vision ⚠️ — AI workload that enables machines to interpret visual information from images and video. Input is always visual (pixels). → Section 4.1
Confidence Score — Numeric value (0-1) indicating how certain a model is about its prediction. Higher = more confident.
Content Filter ⚠️ — Safety mechanism in the Safety System layer that classifies and blocks harmful content based on severity levels. → Section 6.1.3
Copilot ⚠️ — AI assistant integrated into applications that helps users with tasks using plugins and generative AI. Examples: Microsoft 365 Copilot, GitHub Copilot. → Section 6.1.2
D
DALL-E ⚠️ — OpenAI model that generates images from natural language descriptions. Can create, edit, and vary images but CANNOT describe images (that's Computer Vision). → Section 6.1.1
Deep Learning — Machine learning using neural networks with many hidden layers. "Deep" refers to layer count, not quality. Best for unstructured data (images, text, audio). → Section 3.1.4
Document Intelligence ⚠️ — AI workload that extracts STRUCTURED information from documents like invoices and forms. Goes beyond OCR by understanding field relationships. → Section 2.1.3
E
Embeddings ⚠️ — Numerical vector representations of text enabling similarity search, classification, and comparison. Powers semantic search in generative AI applications. → Section 6.1.1
Entity Linking — NLP capability that connects recognized entities to knowledge bases like Wikipedia for additional context.
Entity Recognition (NER) ⚠️ — NLP capability that identifies named entities (people, places, organizations, dates) in text. Different from key phrases (topics) and sentiment (emotion). → Section 5.1.2
F
Face Attributes — Analyzed characteristics of detected faces including age, emotion, glasses, facial hair. Part of Azure AI Face detection (generally available). → Section 4.1.4
Fairness ⚠️ — Responsible AI principle ensuring AI treats all people equitably without discrimination or bias. → Section 2.2.1
Features ⚠️ — Input attributes used to make predictions in machine learning. The "questions" in the exam analogy. Contrast with labels (the "answers"). → Section 3.2.1
G
Generative AI ⚠️ — AI that creates new content (text, images, code) rather than just analyzing existing content. Key distinction: generation vs analysis. → Section 6.1
GPT ⚠️ — Generative Pre-trained Transformer. OpenAI's models for text generation and understanding. The "T" stands for Transformer—if asked about the architecture, answer "transformer." → Section 6.1.1
Grounding ⚠️ — Technique that provides generative AI with specific data context to improve accuracy and reduce hallucinations. Part of the metaprompt layer. → Section 6.1.3
I
Image Classification ⚠️ — Computer vision capability that assigns a single label to an entire image. Returns ONE label, no location information. Contrast with object detection. → Section 4.1.1
Inclusiveness — Responsible AI principle ensuring AI empowers people of all abilities and backgrounds. → Section 2.2.1
K
Key Phrase Extraction ⚠️ — NLP capability that identifies main topics and concepts in text. Finds important TOPICS, not named entities (people, places). → Section 5.1.1
L
Labels ⚠️ — Output values that machine learning models predict; the "correct answers" in supervised learning. Contrast with features (the inputs). → Section 3.2.1
Language Detection — NLP capability that identifies which language text is written in. Part of Azure AI Language service.
Large Language Model (LLM) ⚠️ — Neural network trained on massive text data for language understanding and generation. GPT-4, GPT-3.5 are examples. Based on transformer architecture. → Section 3.1.5
Limited Access ⚠️ — Microsoft policy restricting sensitive AI features (facial recognition, custom neural voice) to approved applications only. Requires application and approval. → Section 4.2.2
Logistic Regression ⚠️ — Despite its name, a CLASSIFICATION algorithm that predicts categories (not numbers). #1 exam trap—memorize this! → Section 3.1.2
M
Machine Learning — AI subset that learns patterns from data to make predictions without explicit programming. Includes supervised, unsupervised, and reinforcement learning. → Section 3.1
Metaprompt ⚠️ — Layer in responsible AI architecture that includes system messages and grounding data. Sets context before user input. → Section 6.1.3
Model Catalog — Azure AI Foundry feature providing access to thousands of AI models from multiple providers (OpenAI, Meta, Mistral, etc.). → Section 6.2.3
Multiclass Classification — Classification with more than two possible outcome categories (e.g., red/blue/green, low/medium/high). → Section 3.1.2
Multiple Linear Regression — Regression using multiple input features to predict one output value. Still regression (predicting numbers), just with more inputs.
N
Named Entity Recognition (NER) — See Entity Recognition. → Section 5.1.2
Natural Language Processing (NLP) ⚠️ — AI workload enabling machines to understand and generate human language. Input is text or speech. → Section 5.1
Neural Network — Computing system with layers of connected nodes; foundation of deep learning. "Deep" = many hidden layers. → Section 3.1.4
O
Object Detection ⚠️ — Computer vision capability that locates AND identifies multiple objects with bounding boxes. Returns labels + coordinates. Contrast with classification (labels only). → Section 4.1.2
OCR (Optical Character Recognition) ⚠️ — Computer vision capability that extracts text from images. It's COMPUTER VISION (image input) even though output is text. Common exam trap! → Section 4.1.3
P
PII Detection — NLP capability that identifies personally identifiable information (names, addresses, SSNs) in text for redaction or compliance.
Pre-built Models ⚠️ — Azure AI models ready to use immediately without training. Contrast with custom models requiring your training data. → Section 1.4.1
Privacy and Security — Responsible AI principle ensuring AI protects personal and sensitive data. → Section 2.2.1
Prompt Engineering ⚠️ — Technique of crafting effective prompts to get desired outputs from generative AI models. Includes system messages, few-shot examples, and output formatting. → Section 6.2.1
R
Regression ⚠️ — Supervised learning technique that predicts continuous numeric values (price, temperature, quantity). Predicting a NUMBER = regression. → Section 3.1.1
Reliability and Safety — Responsible AI principle ensuring AI performs reliably without causing harm. → Section 2.2.1
Responsible AI ⚠️ — Framework of six principles guiding ethical AI: Fairness, Reliability & Safety, Privacy & Security, Inclusiveness, Transparency, Accountability. → Section 2.2
S
Safety System ⚠️ — Layer in responsible AI architecture that applies content filters and abuse monitoring. Catches harmful inputs/outputs. → Section 6.1.3
Semantic Segmentation — Computer vision capability that classifies every pixel in an image. More precise than bounding boxes.
Sentiment Analysis ⚠️ — NLP capability that determines emotional tone (positive/negative/neutral/mixed) of text. Identifies HOW author feels, not WHAT they mention. → Section 5.1.3
Speaker Recognition — Speech capability that identifies distinct speaker voices. Part of Azure AI Speech.
Speech Recognition (Speech-to-Text) ⚠️ — Converting spoken audio to text. Part of Azure AI Speech, not Language. → Section 5.2.2
Speech Synthesis (Text-to-Speech) ⚠️ — Converting text to spoken audio. Part of Azure AI Speech. → Section 5.2.2
Supervised Learning ⚠️ — Machine learning using labeled data (with known correct answers). Labels present = supervised. Includes regression and classification. → Section 1.3.1
System Message ⚠️ — Instructions that set context and behavioral guidelines for generative AI models. Part of metaprompt layer. → Section 6.1.3
T
Tagging — Computer vision feature that associates images with descriptive metadata tags. Part of Azure AI Vision.
Tokenization ⚠️ — Breaking text into individual units (tokens) for processing. Token count affects pricing and context limits in LLMs. → Section 3.1.5
Training Dataset — Data used to teach a machine learning model patterns. Separate from validation and test datasets. → Section 3.2.2
Transformer ⚠️ — Neural network architecture using attention mechanism; powers modern LLMs like GPT. Processes sequences in parallel. The "T" in GPT. → Section 3.1.5
Transparency — Responsible AI principle ensuring users understand AI capabilities, limitations, and how decisions are made. → Section 2.2.1
Translation ⚠️ — Converting text between languages. Azure Translator handles text-to-text; Azure AI Speech handles speech translation. Know the difference! → Section 5.1.6
Transliteration — Converting text to different script (characters) without translating meaning. "東京" → "Tokyo" (same meaning, different script). → Section 5.1.6
U
Unsupervised Learning ⚠️ — Machine learning without labeled data; finds patterns independently. No labels = unsupervised. Clustering is the main example. → Section 1.3.2
V
Validation Dataset — Data used to evaluate model performance during training. Separate from training and test datasets. → Section 3.2.2
Good luck on your AI-900 exam!
This guide was created to help build deep understanding through first principles. For the most current exam objectives, always refer to the official exam page.
