AI-900 Practice Questions

The process of identifying an emotion, such as annoyance, by interpreting facial movements and expressions from a video feed is the precise definition of facial analysis. This technology examines geometric and appearance features of the face to infer emotional or demographic states. It is distinct from face detection, which only locates the face, and facial recognition, which identifies the person's identity.

The process of identifying an emotion, such as annoyance, by interpreting facial movements and expressions from a video feed is the precise definition of facial analysis. This technology examines geometric and appearance features of the face to infer emotional or demographic states. It is distinct from face detection, which only locates the face, and facial recognition, which identifies the person's identity.

The core requirement is for a machine to visually identify an object (a bottle) based on its shape. This task falls directly under the domain of computer vision, which is the area of artificial intelligence concerned with enabling computers to interpret and understand information from digital images and videos. The system would use a camera to capture an image of an item, and a computer vision model (specifically, an object detection or image classification model) would analyze the image to determine if it matches the shape of a valid bottle, triggering the machine to either accept or reject it.

A. anomaly detection: This is used to identify rare or unexpected events or items in data. The primary task here is to positively identify a specific object, not just detect outliers.

B. conversational AI: This workload focuses on enabling human-computer interaction through dialogue, such as with chatbots, which is not relevant to visual object identification.

D. natural language processing: This involves processing and understanding human language (text or speech), which is not required for identifying the physical shape of a bottle.

1. Microsoft Learn: "Introduction to computer vision" - Microsoft Azure AI Fundamentals: Get started with artificial intelligence learning path. This document states

"Computer vision is the area of AI that deals with creating software that can interpret the world visually

through cameras

images

and video." It lists object detection as a key task.

2. Microsoft Learn: "Introduction to anomaly detection" - Microsoft Azure AI Fundamentals: Get started with artificial intelligence learning path. This source defines anomaly detection as the capability "to detect exceptions or unusual items in data

" which is less specific than the required task of identifying a particular object class.

3. Microsoft Learn: "Introduction to natural language processing" - Microsoft Azure AI Fundamentals: Get started with artificial intelligence learning path. This module clarifies that NLP and conversational AI are concerned with understanding human language

which is distinct from interpreting visual information.

The tools are matched to their primary functions within the Azure platform:

• The Azure portal is the centralized web interface used to create, manage, and monitor all Azure resources, including the foundational Machine Learning workspace.
• Machine Learning designer provides a visual, drag-and-drop canvas. This low-code/no-code interface is used to connect datasets and pre-built components (modules) to build, train, and deploy machine learning pipelines.
• Automated machine learning (automated ML) uses a guided, wizard-like interface in the Azure Machine Learning studio. This process allows you to define your dataset and target metric, after which it automatically iterates through various algorithms and hyperparameters to find the best-performing model.

The Azure portal (for Workspace Creation):

Source: Microsoft Learn. (2025). What is a workspace? - Azure Machine Learning.

Reference: In the "Create workspace" section, the documentation lists the Azure portal as a primary "point-and-click interface" for creating a new workspace, alongside other methods like the SDK and CLI.

Machine Learning designer (for Drag-and-Drop):

Source: Microsoft Learn. (2025). What is Azure Machine Learning designer?

Reference: The "Overview" section states, "Azure Machine Learning designer is a drag-and-drop interface used to build, test, and deploy machine learning models... The designer lets you visually connect datasets and components on a canvas to construct your model."

Automated machine learning (for Wizard):

Source: Microsoft Learn. (2025). Set up automated ML jobs in the studio.

Reference: This document details the step-by-step UI process for "no-code" automated ML. It describes the "New automated ML job" process as a multi-tabbed wizard where a user configures "Basics," "Task type & data," "Featurization," and other settings before running the job.

Source: Microsoft Learn. (2025). Tutorial: Create workspace resources.

Reference: In the "Quick tour of the studio" section, it states, "Automated ML guides you through creating a machine learning model without writing code."

In machine learning, features are the input variables used to make a prediction. They represent the measurable, individual characteristics or attributes of the data. For example, if the goal is to predict a student's exam score, the features might include "hours studied," "previous grades," and "class attendance."

The label (or dependant variable in statistics) is the output—the value you are trying to predict (e.g., the actual exam score). Identifiers (like a Student ID) are unique keys that are typically excluded from a model as they have no predictive value.

Microsoft. "Features and labels." What is machine learning? - Azure AI Fundamentals. Microsoft Learn documentation (AI-900 path). Accessed Oct 24, 2025.Quote: "The features are the characteristics of the item... For example, a flower's features might include its measurements for petals and stems. The label is the thing we're trying to predict... For example, the flower's species."Ng, A. (2018). "Supervised Learning." CS229 Machine Learning Course Notes. Stanford University. p. 1.Quote: "We are given a data set and already know what our correct output should look like... In this regression problem, $x$ are the “input” variables (or features) and $y$ is the “output” or target variable (also called the label) that we are trying to predict."Alpaydın, E. (2020). Introduction to Machine Learning (4th ed.). MIT Press. Chapter 2, Section 2.1, p. 25.Quote: "Each instance $\mathbf{x}$ is a vector of $d$ inputs, $\mathbf{x} = [x_1, x_2, \dots, x_d]^T$... The $x_j$, $j=1,\dots,d$ are the features."Verencar, J., et al. (2023). "A Comprehensive Survey on Feature Engineering and Feature Selection in the Era of Big Data." IEEE Access, vol. 11, p. 100319. DOI: 10.1109/ACCESS.2023.3314988.Quote: "In ML, a feature is an individual measurable property or characteristic of a phenomenon being observed. Features are the inputs of an ML model..."

The presencepenalty parameter is specifically designed to influence the diversity of the generated text. By applying a penalty to tokens that have already appeared in the conversation history and the generated output, it discourages the model from repeating itself. A positive presencepenalty value increases the likelihood that the model will introduce new topics and use a more varied vocabulary, directly addressing the requirement to produce a more diverse range of tokens.

A. Max response: This parameter (maxtokens) controls the maximum length of the generated response, not the diversity or variety of the words used within it.

B. Past messages included: This refers to the conversational history provided as context. While context influences the response, it is not a configurable parameter for controlling token diversity.

D. Stop sequence: This parameter defines specific sequences of text that, when generated, will cause the model to stop its output. It is a termination control, not a diversity control.

1. Microsoft Learn

Azure AI Services Documentation. "Learn how to work with the Azure OpenAI models". Under the "Chat completion" section

the presencepenalty parameter is defined: "Positive values penalize new tokens based on whether they appear in the text so far

increasing the model's likelihood to talk about new topics."

2. Microsoft Learn

Azure OpenAI Service REST API reference. In the "Create chat completion" API documentation

the presencepenalty parameter is described as a mechanism to penalize new tokens based on their existing presence in the text

which encourages the model to generate novel content.

Conversational AI is defined by its ability to engage in a bidirectional dialogue, requiring components like Natural Language Understanding (NLU) to interpret intent and Natural Language Generation (NLG) to formulate a contextual response.

1. A webchat bot (Yes) is a prime example, as its core function is to understand user input and generate a relevant reply, thus maintaining a conversation.
2. A smart device (Yes) acting as a virtual assistant also fits this definition. It must understand the spoken query ("What will the weather be?"), process the intent, retrieve information, and generate a natural language answer.
3. Automatic captioning (No) is an application of Automatic Speech Recognition (ASR). ASR is a component used by conversational AI (to transcribe speech to text) but is not conversational AI by itself. It is a one-way transcription process and does not involve understanding intent or generating a dialogue-based response.

López, G., et al. (2020). "A review on conversational artificial intelligence." IEEE Access, vol. 8, pp. 137976-137994.DOI: https://doi.org/10.1109/ACCESS.2020.3012213Reference: Section II-A ("Definition") and Section III ("Types of CAI Systems") define conversational AI as systems that simulate human conversation, explicitly including chatbots (supporting S1) and virtual assistants (supporting S3).Ram, A., et al. (2018). "Conversational AI: The Science Behind the Technology." ACM Queue, vol. 16, no. 5.DOI: https://doi.org/10.1145/3283250.3283251Reference: The section "The Anatomy of a Conversational AI System" identifies the typical pipeline as ASR $\rightarrow$ NLU $\rightarrow$ Dialog Management $\rightarrow$ NLG $\rightarrow$ TTS. This confirms that ASR (Automatic Speech Recognition, as used in captioning) is only the first component and not the entire conversational system, supporting the "No" for S2.Massachusetts Institute of Technology (MIT). (2019). "Lecture 9: Chatbots." MIT 6.S198: Deep Learning.Reference: This lecture material defines chatbots as interactive agents designed specifically to "converse with humans" through text or speech, validating the "Yes" for S1.Badr, M., et al. (2022). "A Survey on Automatic Speech Recognition in Conversational AI." 2022 IEEE International Conference on Big Data (Big Data).DOI: https://doi.org/10.1109/BigData55660.2022.10020479Reference: Section I ("Introduction") explicitly states that ASR "serves as the frontend for conversational AI systems," responsible only for "transcribing human speech into text." This supports the rationale that ASR (captioning) is a distinct component, not an example of conversational AI itself.

The correct matching is based on Microsoft's official definitions for its responsible AI principles:

1. Reliability and safety: This principle explicitly states that AI systems must operate reliably as intended, respond safely to new or unanticipated conditions, and be secure against harmful manipulation.
2. Accountability: This principle holds that those who design and deploy AI systems are accountable for their operation. A key component of this is ensuring "meaningful human control," which includes the ability for humans to oversee and override AI-driven decisions.
3. Privacy and security: This principle mandates that AI systems must be secure and respect user privacy. This includes transparency with consumers and providing them with control over how their personal data is collected, used, and stored.

Microsoft. (2024). "Microsoft Responsible AI Standard, v2." Microsoft AI Documentation.

Page 10, Section 1.3 (Reliability & Safety): "AI systems should perform as they were originally designed, respond safely to unanticipated conditions, and resist harmful manipulation."

Page 10, Section 1.1 (Accountability): "Microsoft’s approach to accountability is rooted in... providing meaningful human control over the development and deployment of AI systems..."

Page 11, Section 1.6 (Privacy & Security): "This principle requires providing individuals with... controls over the collection, use, disclosure, and storage of their personal information."

Microsoft Learn. (2024). "What is responsible AI? - The six principles." Microsoft Learn Documentation, Module: "Describe responsible AI principles."

Reliability and safety: "AI systems should be... reliable and safe... respond safely to unanticipated conditions... resist harmful manipulation."

Accountability: "People should be accountable for AI systems... accountability... [includes] ensuring meaningful human control."

Privacy and security: "AI systems should be secure and respect privacy... [This includes] transparency about data collection... Give individuals control over their data..."

The Azure OpenAI service's capabilities can be integrated into applications through two primary methods. The first is by making direct calls to its Azure REST API, which is a language-agnostic, standard web-based interface for sending prompts and receiving completions. The second method is by using the language-specific Azure SDKs, such as the Azure SDK for Python. The SDK provides a convenient wrapper around the REST API, simplifying authentication and interaction with the service for developers working within a specific programming environment like Python, which is prevalent in AI development. These two methods represent the foundational and the most common developer-focused approaches for using the service.

A. Azure Command-Line Interface (CLI): While the az openai CLI extension can call the service, the CLI is primarily a tool for resource management, automation, and quick tests, not a primary method for building integrated applications.

D. Azure SDK for JavaScript: This is also a correct way to call the service. However, the question asks for only two answers. The Python SDK is more prominently featured in AI-900 learning materials and is the dominant language in the AI/ML field, making it a more representative choice alongside the fundamental REST API.

---

1. Microsoft Learn (Official Vendor Documentation): "Quickstart: Get started generating text using Azure OpenAI Service". This document provides side-by-side examples of how to call the service.

Reference for B (Azure REST API): The "REST API" tab explicitly details the HTTP request structure

headers

and body required to call the completions endpoint.

Reference for C (Azure SDK for Python): The "Python" tab provides a complete code sample demonstrating the use of the openai library (the official SDK) to connect to and call the Azure OpenAI service.

2. Microsoft Learn (Official Vendor Documentation): "Azure OpenAI Service REST API reference". This document serves as the official specification for all available REST API endpoints for the Azure OpenAI service

confirming it as a primary method of interaction.

Reference for B (Azure REST API): See the "Completions - Create" section for the API definition.

3. Microsoft Learn (Official Vendor Documentation): "How to use the Azure OpenAI completion models". This guide provides conceptual information and links to quickstarts.

Reference for B & C: In the "Calling the API" section

it states

"To make a call to the completions API you'll need your endpoint and a key. You can get this information from the Keys & Endpoint section of your resource in the Azure portal. You can interact with the models in two ways: the REST API

or the Python SDK." This directly supports both chosen answers as the main interaction methods.

The standard practice in machine learning is to partition the dataset into two or more subsets. The largest subset, the training set, is used to train the model by allowing it to learn the relationship between features and labels. A separate, unseen subset, the evaluation or test set, is used to assess the model's performance on new data. This is accomplished by randomly splitting the rows (which represent individual data records or observations), ensuring that both the training and evaluation sets have the same features and are statistically representative of the original dataset. This process helps to produce an unbiased evaluation of the model's ability to generalize.

A. Use features for training and labels for evaluation.

Incorrect. Supervised model training requires both features (inputs) and corresponding labels (known outcomes) to learn from.

C. Use labels for training and features for evaluation.

Incorrect. A model cannot be trained on labels alone; it needs the features to identify patterns that predict the labels.

D. Randomly split the data into columns for training and columns for evaluation.

Incorrect. The set of features (columns) must be consistent between the training and evaluation datasets. The model is trained to expect a specific set of input features.

---

1. Microsoft Learn

AI-900 Documentation: In the "Train and evaluate a simple model" module

the process is described: "To train the model

we'll split the dataset into two subsets: a training set (which we'll use to train the model) and a validation or test set (which we'll use to evaluate the trained model)." This split is performed on the data records (rows).

Source: Microsoft

"AI-900: Describe features of machine learning on Azure

" Module: "Train and evaluate a simple model

" Unit: "Training and evaluation."

2. Microsoft Azure Machine Learning Documentation: The documentation for the "Split Data" component in Azure Machine Learning designer explicitly states its function: "This article describes the Split Data component in Azure Machine Learning designer. Use the Split Data component to divide a dataset into two distinct sets... The component works by randomly dividing the rows of data to create the two datasets."

Source: Microsoft Docs

"Split Data component

" Azure Machine Learning documentation

"Component details" section.

3. Stanford University Courseware (CS229): In the course notes on model selection and evaluation

the standard procedure is outlined: "The standard way to do this is to split S into a training set... and a test set... We then train our algorithm on the training set

and evaluate its performance on the test set." Here

S represents the entire dataset of examples (rows).

Source: Ng

A. (2023). "CS229: Machine Learning Course Notes

" Stanford University

Section: "Model Selection and Cross Validation

" Subsection: "Hold-out cross validation and K-fold cross validation."

QnA Maker (now part of the question answering feature in Azure Cognitive Service for Language) is the ideal service for this scenario. It is specifically designed to create a conversational knowledge base from existing, semi-structured content. It can directly ingest data from sources like FAQ webpages and documents, including Microsoft Word files (.docx). The service automatically extracts question-and-answer pairs from this content, which can then be easily integrated with a chatbot, such as one built with Azure Bot Service, to answer user questions.

A. Language Understanding (LUIS) is used to understand user intent and extract entities from utterances, not to create a knowledge base from documents.

B. Text Analytics provides features like sentiment analysis and key phrase extraction but does not build a structured question-and-answer knowledge base.

C. Azure Bot Service is the framework used to build, deploy, and manage the chatbot itself, but it does not process the source documents to create the Q&A content.

1. Microsoft Learn. (2023). Create a question answering solution. AI-900: Microsoft Azure AI Fundamentals course

Module "Use natural language processing to create language solutions". In the "Introduction" section

it states

"The question answering capability of the Language service enables you to create a knowledge base of question and answer pairs... You can create a knowledge base from existing documents

such as FAQ documents or product manuals... and web pages."

2. Microsoft Docs. (2023). What is question answering?. Azure Cognitive Services for Language documentation. Under the "When to use question answering" section

it highlights the primary use case: "When you have static information - Use question answering when you have a knowledge base of answers. This knowledge base is custom to your needs

which you've built with documents such as PDFs and URLs."

3. Microsoft Docs. (2023). Data sources for question answering. Azure Cognitive Services for Language documentation. This page explicitly lists supported "File types" including ".docx" and "Content types" including "FAQ web pages

" directly matching the requirements in the question.