Azure AI Fundamentals AI-900 练习

AI-900 focuses on describing AI workloads, machine learning principles, computer vision, natural language processing, generative AI, and responsible AI. It does not require deep coding, VM management, or database design.

Azure AI services expose prebuilt models for vision, language, speech, and decision making through REST APIs and SDKs. Virtual machines, storage, and DevOps serve different purposes.

An AI workload applies models to data so the system can understand, predict, or generate results. Backups, load balancers, and updates are infrastructure operations, not AI workloads.

OCR reads text from images and documents. Image classification labels whole images, segmentation labels pixels, and sentiment analysis judges text tone.

Generative AI produces new text, images, audio, or other content from learned patterns. Regression predicts numbers, clustering groups data, and NER extracts entities.

Transparency means people understand how the system works, what it can do, and where it can fail. Fairness addresses bias, reliability addresses dependable behavior, and inclusiveness addresses access.

Fairness means AI systems should avoid unfair bias and treat groups consistently. Accountability assigns responsibility, privacy protects data, and transparency explains behavior.

Privacy and security require protecting sensitive data, controlling access, and following data governance rules. The other principles cover access, explanation, and bias.

Machine learning principles cover supervised tasks such as regression and classification plus unsupervised tasks such as clustering. The other domains focus on images, language, and generation.

Computer vision, NLP, and generative AI are AI workload categories. Azure DNS is a networking service, not an AI workload.

Supervised learning uses inputs with known labels so the model can learn the relationship. Unsupervised learning finds patterns without labels, and clustering is an unsupervised task.

Regression predicts continuous numeric values such as price or temperature. Classification predicts categories, clustering groups data, and NER extracts entities from text.

Classification assigns items to predefined categories. Regression predicts numbers, clustering finds groups, and key phrase extraction identifies important phrases.

Clustering is an unsupervised task that groups similar data points. Classification and regression are supervised, and object detection works with images.

Features are the input variables the model uses to make predictions. The label is the target output, and deployment or scheduling are operational concerns.

A label is the known correct output for each training example. Features are inputs, and regions or files are storage details.

Overfitting happens when a model memorizes training details instead of general patterns, so it fails on new data. Underfitting is too simple, and speed or reading issues are different.

Azure Machine Learning provides workspaces, experiments, pipelines, and endpoints for the ML lifecycle. App Service hosts web apps, Cosmos DB stores data, and Front Door routes traffic.

AutoML tests candidate algorithms and settings to find a good model for your data. It does not delete resources, translate output, or back up databases.

Regression and classification are supervised tasks with labeled outputs. Clustering is unsupervised, and key phrase extraction is an NLP task.

Image classification labels the whole image, such as dog or cat. Object detection finds objects and boxes, segmentation labels pixels, and OCR extracts text.

Object detection locates objects and returns bounding boxes plus class labels. Classification labels the whole image, face verification compares identities, and translation works on text.

Semantic segmentation assigns a class to each pixel, producing a detailed mask. Classification labels the whole image, and the other options are NLP tasks.

Azure AI Vision includes OCR capabilities for printed and handwritten text. Translator converts language, sentiment analysis judges tone, and Speech processes audio.

Face recognition verifies or identifies people by comparing face features. Face detection only finds face locations, and the other options do not compare identities.

Face detection locates faces and may return attributes such as pose or emotion. Face recognition identifies who a person is, OCR reads text, and sentiment analysis judges tone.

Azure AI Vision performs image analysis, including tags, descriptions, and categories. Translator works with language, OpenAI provides generative models, and Monitor observes systems.

Document Intelligence, formerly Form Recognizer, extracts fields, tables, and key-value pairs from documents. Object detection finds objects in images, and the other services are not document analyzers.

OCR reads the characters in an image, which makes it suitable for license plates. Classification labels whole images, sentiment works on text, and speech recognition works on audio.

Classification returns one label, object detection returns boxes, segmentation labels each pixel, and OCR returns extracted text.

Sentiment analysis classifies the emotional tone of text. Key phrase extraction finds topics, NER finds entities, and speech synthesis creates spoken audio.

Named entity recognition identifies entities such as people, locations, and organizations. Language detection identifies the language, summarization shortens text, and captioning describes images.

Key phrase extraction returns the most important phrases in a document. Entity linking connects mentions to knowledge base entries, and the other options are speech or vision tasks.

Azure AI Translator converts text between supported languages. Vision processes images, Speech processes audio, and Policy enforces governance rules.

Azure AI Speech includes speech to text, text to speech, translation, and speaker recognition. Translator handles written text, Vision handles images, and Machine Learning builds models.

An intent is the user goal, such as booking a flight or checking weather. Entities provide details, and scores, language codes, or tables are not intents.

Entities capture details such as dates, cities, or amounts that help fulfill the intent. Language, response format, and model type are different concepts.

Language detection returns the language of the input text. Sentiment judges tone, OCR reads images, and text to speech creates audio.

Custom question answering matches user questions to answers from provided content. The other capabilities identify phrases, entities, or language.

Sentiment judges tone, NER extracts entities, Translator converts languages, and Speech processes audio.

Generative AI produces new text, images, audio, or video by learning patterns from training data. Classification labels data, and storage or monitoring are not generative.

LLMs learn language patterns from massive text datasets and can generate completions, summarize, translate, and answer questions. They are not databases, firewalls, or VM images.

A prompt is the user-provided instruction or context. The generated result is a completion, labels are used in training, and regions are deployment locations.

A completion is the model-generated output that follows the prompt. It is not a transaction, metric, or request.

Tokens are the text chunks, such as words or parts of words, that a model reads and generates. They are not billing credits, processors, or rows.

Higher temperature increases randomness and creativity, while lower temperature makes output more focused and predictable. Max tokens limits length, and the other values are training concepts.

Embeddings map text to numeric vectors so similar meanings are close together, which supports search and similarity. They are not images, passwords, or cables.

Azure OpenAI provides models such as GPT for text and DALL-E for images. SQL Database stores data, Front Door routes traffic, and Backup protects data.

RAG retrieves relevant documents and includes them in the prompt so the model can answer with evidence. It is not training without data, encryption, or compression.

A prompt is the input, a completion is the output, temperature controls randomness, and embeddings represent meaning numerically.

Fairness focuses on avoiding bias and treating groups equitably. Inclusiveness is about access, transparency about explanation, and accountability about responsibility.

Explainability tools and documentation show how inputs affect outputs so decisions can be understood. The other options do not improve understanding.

Human-in-the-loop keeps people involved in reviewing, correcting, or approving AI outputs, which improves safety and accountability.

The Responsible AI dashboard provides error analysis, model explanations, fairness assessments, and counterfactual insights. DNS, app logs, and VNet peering are unrelated.

Transparency requires clear disclosure and explanation of what the system can and cannot do. Hiding behavior, unlimited storage, and removing controls are opposite practices.

Access control, encryption, and governance policies keep sensitive data safe. Publishing data, removing logs, and disabling authentication increase risk.

Reliability and safety mean predictable, dependable behavior with safe handling of failures. The other options describe risky or incorrect behavior.

Inclusiveness calls for AI that works for people with varied abilities, backgrounds, and needs. Fairness addresses bias, privacy protects data, and transparency explains behavior.

Accountability means individuals and organizations own the decisions and outcomes of AI systems. The other principles address bias, behavior, and access.

Content filters, human review, clear labeling, and bias testing all reduce generative AI harms. Together they improve safety, transparency, and fairness.