Azure AI-900 cheat sheet

AI workload families in scope: machine learning, computer vision, natural language processing, document intelligence, knowledge mining, and generative AI.

The six Responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability.

Fairness means equitable treatment across groups; biased or unrepresentative training data is the usual root cause of unfair model outcomes.

Reliability and safety means the system operates consistently and does not cause harm, which requires testing, monitoring, and fail-safe design.

Transparency means stakeholders can understand the system's purpose, data, and decision logic; explainability is the mechanism that supports it.

Accountability means people and organizations remain answerable for AI outcomes, implemented through governance boards, approvals, and audit logs.

Inclusiveness means the solution works for all users, including people with disabilities; speech, vision, and language features support accessibility.

Privacy and security means personal data is collected lawfully, protected in transit and at rest, and used only for its stated purpose.

Scalability and cost-optimization are engineering qualities, not one of the six Responsible AI principles.

Supervised learning uses labeled data; its two core tasks are regression (numeric output) and classification (categorical output).

Unsupervised learning uses unlabeled data; the main task in scope is clustering, which groups similar records by feature similarity.

Regression is evaluated with MAE, MSE, RMSE, and R-squared; classification is evaluated with accuracy, precision, recall, and F1 score.

Precision = true positives / (true positives + false positives); recall = true positives / (true positives + false negatives).

Overfitting shows as strong training performance with weak validation or test performance; remedies include more data, a simpler model, or regularization.

Data is normally split into training, validation, and test sets; the test set estimates performance on unseen data.

An Azure Machine Learning workspace is the top-level resource that holds compute targets, datastores, datasets, models, and pipelines.

Automated ML selects algorithms and hyperparameters automatically, while the Azure Machine Learning designer provides a drag-and-drop pipeline canvas.

Feature engineering creates or transforms input columns (for example normalizing values) so the model can learn the relationship more effectively.

Image classification assigns one or more labels to an entire image; object detection returns labels plus bounding boxes locating each object.

Semantic segmentation classifies every pixel into a category, producing a mask that shows the exact shape of each region.

OCR extracts printed or handwritten text from images and PDF files and converts it into machine-readable characters.

Prebuilt Azure AI Vision capabilities include image tagging, captioning, object detection, face detection, OCR/Read, and spatial analysis.

Custom Vision trains image classification or object detection models on your own labeled images and is a separate service from Azure AI Vision.

Azure AI Document Intelligence (formerly Form Recognizer) extracts fields, key-value pairs, and tables from forms, receipts, and invoices.

Face detection returns face locations, verification compares two faces, and identification matches a face to a known person; these are limited-access features.

Object detection accuracy is summarized with mean average precision (mAP), computed across classes and intersection-over-union thresholds.

A bounding box is defined by its top-left x and y coordinates plus a width and a height.

Azure AI Language capabilities include language detection, sentiment analysis, key phrase extraction, named entity recognition, summarization, PII detection, and question answering.

Sentiment analysis returns positive, neutral, or negative labels with confidence scores between 0 and 1 at both document and sentence level.

Named entity recognition classifies entities such as person, location, organization, datetime, and quantity; PII detection flags personal identifiers for redaction.

Common text preprocessing steps are tokenization (splitting text into tokens), stop-word removal, and stemming or lemmatization.

TF-IDF scores a term by its frequency in a document, weighted down by the number of documents that contain it.

Azure AI Translator translates text between languages; Azure AI Speech provides speech-to-text, text-to-speech, and speech translation.

Conversational Language Understanding (CLU) extracts intents and entities from user utterances so a bot can decide what to do next.

Question answering in Azure AI Language returns the best answer from a curated knowledge base instead of generating free-form text.

Generative models create new content such as text, images, or code by learning patterns in training data; text models work by predicting the next token.

A prompt typically contains a system message (behavior rules) and a user message (the task); the model responds with a completion.

Tokens are chunks of text, roughly four characters of English each; max tokens and model choice drive cost and latency.

Temperature and top_p control randomness: lower temperature gives more deterministic output, higher temperature gives more varied output.

Retrieval augmented generation (RAG) retrieves grounding data from a search index and injects it into the prompt, reducing hallucination without retraining.

Fine-tuning retrains a base model on your own examples to change style or format, while RAG supplies current facts at query time.

Azure OpenAI Service hosts OpenAI models (GPT, DALL-E, embeddings) inside Azure with private networking, compliance coverage, and content filters.

Content filtering in Azure OpenAI scores prompts and completions against hate, sexual, violence, and self-harm categories and can block them.

Hallucination is fluent but unsupported or fabricated output; mitigations include grounding with RAG, clear system instructions, and human review.