AI-900 cheat sheet
Decision rules for choosing the right Azure AI service, memory aids, and the reference tables I reviewed on exam day.
Decision rules
Identifying WHAT something is? → Image Classification
Finding WHERE something is? → Object Detection
Same person? (1:1) → Face Verification
Who is this person? (1:many) → Face Identification
Age, glasses, orientation? → Face Detection (NOT recognition)
Extract text from image? → Vision OCR
Translate text? → Translator service (NOT Vision)
Moderate content? → Content Safety (NOT Vision)
Understanding intent? → Language service (CLU)
Speech to text? → Speech service
Variable workload, pay-per-use? → Standard deployment
Large batch, async? → Batch deployment
Mission-critical, low-latency? → Provisioned deployment
Who is responsible for AI? → PEOPLE (not the AI)
Security and privacy principle? → Privacy & Security (NOT Reliability & Safety)
Memory Aids
Services
Vision = Visual analysis (see images) Language = Linguistic analysis (understand text) Speech = Sound (audio ↔ text) Translator = Transform languages
Face Operations
Verify = Vs comparison (1 vs 1) Identify = Individual lookup (1 vs many)
Detection vs Recognition
Detection = Describe attributes (age, glasses) Recognition = Recognize identity (verify, identify)
Responsible AI
Fairness → Fair treatment Reliability → Runs correctly Privacy → Protects data Inclusiveness → Includes everyone Transparency → Transparent/understandable Accountability → Adults (people) responsible
Confidence Scores
0 to 1 (NOT -1 to 1) Think: 0% to 100% confidence
Fine-tuning Format
JSONL = Just right CSV = Can't use
Transformer Speed
Parallel = Powerful (fast) Sequential = Slow (RNNs)
AutoML
Classification Regression Time-series Clustering (only K-means exists)
Reference tables
Responsible AI Principles - One-Liner Summary
| Principle | One-Liner |
|---|---|
| Fairness | No unfair advantage/disadvantage to any group |
| Reliability & Safety | Performs reliably, tested, safe operation |
| Privacy & Security | Secure systems, data protection, respects privacy |
| Inclusiveness | Benefits all people, accessible, empowering |
| Transparency | Users understand how AI works, limitations, capabilities |
| Accountability | People (designers, developers) are responsible |
Azure AI Services Capability Matrix
| Capability | Vision | Face | Language | Speech | Translator | Content Safety | Custom Vision |
|---|---|---|---|---|---|---|---|
| Image analysis | |||||||
| OCR | |||||||
| Face detection | |||||||
| Face recognition | |||||||
| Content moderation | |||||||
| Sentiment analysis | |||||||
| NER | |||||||
| Speech-to-text | |||||||
| Translation | |||||||
| Train custom models |
Computer Vision Decision Tree
Question Type → Answer
"What type/category is this?" → Image Classification
"Where is it located?" / "bounding boxes" → Object Detection
"Classify each pixel" → Semantic Segmentation
"Extract text from image" → OCR
"Generate description of image" → Image Analysis
"Train custom categories" → Custom Vision
"Use pre-trained model" → Azure AI VisionFace Services Decision Tree
Question Type → Answer
"Age, glasses, orientation?" → Face Detection
"Same person? (1:1)" → Face Verify
"Who is this? (1:many)" → Face Identify
"Find similar faces from list" → Find Similar
"Group faces by similarity" → Group
"Can emotion be detected?" → NO (removed for ethical reasons)ML Type Decision Tree
Question Type → Answer
"Predict a number" → Regression
"Predict a category" → Classification
"Group unlabeled data" → Clustering
"AutoML supports?" → Classification, Regression, Time-series (NOT Clustering)