View Full Microsoft AI-901 Exam Dumps and Practice Test Dumps

 

Question 221. Which Azure Machine Learning feature can be used to track experiments and training runs?

1. Azure Machine Learning jobs
2. Azure DNS
3. Azure Firewall
4. Azure Storage Explorer

Correct Answer: 1. Azure Machine Learning jobs

Explanation:

Azure Machine Learning jobs provide a way to run and track machine-learning activities such as training, evaluation, and data-processing tasks. Tracking jobs can help developers compare different experiments and review information such as metrics, parameters, outputs, and execution details. This is useful when a team needs to understand which training configuration produced a particular model and whether subsequent experiments improved performance. Azure DNS and Azure Firewall serve networking and security purposes, while Azure Storage Explorer is primarily a tool for interacting with storage resources. Therefore, Azure Machine Learning jobs are appropriate for managing and tracking machine-learning experiment runs.

Question 222. What is the purpose of model evaluation before deployment?

1. To eliminate the need for training
2. To determine whether the model meets defined performance requirements
3. To create network addresses
4. To store passwords

Correct Answer: 2. To determine whether the model meets defined performance requirements

Explanation:

Model evaluation helps determine whether a trained model performs adequately for its intended purpose before it is deployed. Developers can use appropriate evaluation datasets and metrics to measure characteristics such as accuracy, precision, recall, error, or other application-specific criteria. Evaluation can also reveal whether the model has problems with generalization, fairness, or other important requirements. A model should not be considered ready simply because it completed training successfully. Network addressing and password storage are unrelated activities. Therefore, evaluation provides evidence that the model satisfies the technical and business requirements established for its intended use.

Question 223. Which metric is commonly used to evaluate the error of a regression model by squaring individual errors?

1. Recall
2. Precision
3. Mean squared error
4. Accuracy

Correct Answer: 3. Mean squared error

Explanation:

Mean squared error, or MSE, calculates the average of the squared differences between predicted values and actual values. It is commonly used to evaluate regression models because larger errors receive greater mathematical weight after being squared. This can make MSE particularly useful when large prediction errors are especially important to identify. For example, when evaluating a model that predicts numerical demand, a few very large errors can contribute significantly to the MSE. Accuracy, precision, and recall are commonly associated with classification problems. Therefore, when the requirement specifically mentions squaring prediction errors in a regression task, mean squared error is the appropriate metric.

Question 224. What does the F1 score combine in a classification model?

1. Precision and recall
2. Accuracy and storage
3. Training time and memory
4. Features and endpoints

Correct Answer: 1. Precision and recall

Explanation:

The F1 score combines precision and recall into a single metric by calculating their harmonic mean. It can be useful when both types of performance are important and a balance between identifying positive cases and avoiding incorrect positive predictions is required. A model with high precision but very low recall may not have a strong F1 score, and the same applies in the opposite situation. F1 can therefore provide a useful summary when evaluating classification performance. Training time, storage, features, and endpoints are separate concepts and are not components of the F1 score calculation.

Question 225. Which Azure Machine Learning capability helps provide scalable compute for machine-learning workloads?

1. Azure AI Translator
2. Compute clusters
3. Azure AI Vision
4. Azure DNS

Correct Answer: 2. Compute clusters

Explanation:

Azure Machine Learning compute clusters provide scalable managed compute resources for machine-learning workloads. They can be used for tasks such as model training and other computational jobs where organizations need more processing capacity than a local computer can provide. Compute clusters can scale according to workload requirements, making them useful for repeated or resource-intensive training activities. Azure AI Translator is designed for language translation, Azure AI Vision handles computer vision tasks, and Azure DNS provides name-resolution services. Therefore, compute clusters are the Azure Machine Learning resource that directly addresses the requirement for scalable compute resources.

Question 226. What is model drift?

1. A change in a model’s performance or relationship with real-world data over time
2. A method for compressing models
3. A technique for translating documents
4. A way to create database tables

Correct Answer: 1. A change in a model’s performance or relationship with real-world data over time

Explanation:

Model drift refers to changes that can cause a deployed machine-learning model to become less effective as real-world conditions evolve. Customer behavior, market conditions, user preferences, or other environmental factors may change after the model has been trained. As a result, relationships learned from historical data may no longer represent current conditions accurately. Monitoring helps organizations identify changes in model performance and determine whether investigation or retraining is necessary. Model drift is not related to file compression, document translation, or database table creation. It is an important consideration for maintaining model quality after deployment.

Question 227. Which type of data change occurs when the statistical distribution of model inputs changes over time?

1. Data drift
2. Text summarization
3. Object detection
4. Model registration

Correct Answer: 1. Data drift

Explanation:

Data drift occurs when the distribution or characteristics of input data change over time compared with the data used during model development. For example, a model trained on historical customer behavior may receive substantially different customer behavior after a major change in market conditions. Even if the model itself has not changed, its inputs may no longer resemble the data on which it was trained. Monitoring data distributions can help identify this situation. Text summarization and object detection are AI capabilities, while model registration concerns managing model versions. Therefore, a changing statistical distribution of model inputs is an example of data drift.

Question 228. Which practice helps protect sensitive information used by an AI application?

1. Collecting every possible personal detail
2. Publishing credentials in prompts
3. Applying data minimization and appropriate access controls
4. Giving every user administrator permissions

Correct Answer: 3. Applying data minimization and appropriate access controls

Explanation:

Data minimization means collecting and processing only the information that is necessary for the intended purpose of an AI application. Combined with appropriate access controls, it can reduce the likelihood and impact of unauthorized access to sensitive information. For example, an application that only needs a customer’s region may not need to store unrelated personal details. Access permissions should also be limited according to legitimate requirements. Publishing credentials or giving every user administrator permissions creates significant security risks. Therefore, data minimization together with suitable access controls is an important practice for protecting sensitive information in AI solutions.

Question 229. Which responsible AI principle focuses on assigning responsibility for decisions and outcomes?

1. Accountability
2. Image classification
3. Regression
4. Vector search

Correct Answer: 1. Accountability

Explanation:

Accountability means that people and organizations remain responsible for the design, development, deployment, operation, and outcomes of AI systems. An AI system should not be treated as independently responsible for important decisions. Organizations need appropriate processes for governance, monitoring, review, and responding to problems. For example, if an AI application produces an unexpected result, there should be a clear process for investigating the issue and determining appropriate corrective action. Transparency helps explain an AI system, while fairness addresses equitable treatment. Regression and vector search are technical methods. Therefore, accountability is the responsible AI principle associated with responsibility for AI system outcomes.

Question 230. What is the main purpose of access control in an AI solution?

1. To increase model accuracy automatically
2. To determine which users or services can access resources
3. To generate synthetic speech
4. To classify images

Correct Answer: 2. To determine which users or services can access resources

Explanation:

Access control determines which users, applications, or services are permitted to access specific resources. In an AI solution, resources might include datasets, models, endpoints, storage accounts, search indexes, or other services. Applying the principle of least privilege means providing only the permissions necessary for a user or application to perform its required tasks. This reduces the potential impact of compromised credentials or unauthorized activity. Access control does not directly increase model accuracy, generate speech, or classify images. Those are separate capabilities. Therefore, the primary purpose of access control is to manage who or what can access protected resources.

Question 231. Which Azure AI service can analyze the sentiment expressed in written customer feedback?

1. Azure AI Vision
2. Azure AI Language
3. Azure AI Speech
4. Azure AI Translator only

Correct Answer: 2. Azure AI Language

Explanation:

Azure AI Language provides natural language processing capabilities that can analyze written content, including sentiment analysis. Sentiment analysis can help identify whether text expresses positive, negative, neutral, or mixed sentiment depending on the service capabilities and input. This can be useful for analyzing customer reviews, surveys, support tickets, and other textual feedback. Azure AI Vision focuses on visual information, Azure AI Speech provides speech-related capabilities, and Azure AI Translator focuses on language translation. Therefore, Azure AI Language is the appropriate Azure service when the primary requirement is to analyze sentiment in written customer feedback.

Question 232. Which capability can summarize a long document into a shorter representation of its main ideas?

1. Text summarization
2. Object detection
3. Image segmentation
4. Speech synthesis

Correct Answer: 1. Text summarization

Explanation:

Text summarization is a natural language processing capability that produces a shorter representation of a longer text while attempting to preserve its important information. It can be useful for processing reports, articles, customer feedback, meeting notes, or other large documents. Instead of requiring a user to read the entire source document, a summarization system can provide a concise overview of key points. Object detection and image segmentation work with visual information, while speech synthesis converts written text into spoken audio. Therefore, text summarization is the appropriate capability when the requirement is to reduce a long document into a shorter version containing its main ideas.

Question 233. Which Azure AI capability can extract text from a photograph of a printed sign?

1. OCR
2. Regression
3. Clustering
4. Recall

Correct Answer: 1. OCR

Explanation:

Optical Character Recognition, or OCR, extracts visible text from images such as photographs, scanned pages, and printed signs. An AI application can use OCR to convert visual characters into machine-readable text that can then be searched, stored, translated, or processed by other services. For example, a mobile application could capture a photograph of a sign and extract the text for later use. Regression and clustering are machine-learning techniques, while recall is an evaluation metric. None of those directly perform text extraction from images. Therefore, OCR is the appropriate capability for extracting printed text from a photograph.

Question 234. What is the main benefit of vector search in an AI application?

1. It can retrieve information based on semantic similarity
2. It creates physical network cables
3. It encrypts every document automatically
4. It replaces all language models

Correct Answer: 1. It can retrieve information based on semantic similarity

Explanation:

Vector search uses numerical representations of content, commonly called embeddings, to retrieve information based on similarity in meaning. This can be useful when a user’s query does not use exactly the same words as the relevant document. For example, a user might ask about “ways to secure an account,” while a document discusses “identity protection practices.” Vector search can compare the embeddings of the query and documents to identify semantically related content. It does not replace language models or automatically encrypt documents. Its primary value is improving retrieval by using relationships represented in vector space.

Question 235. Which scenario is most suitable for an online machine-learning endpoint?

1. Processing millions of records overnight without immediate responses
2. Providing a prediction immediately after each application request
3. Archiving old model files
4. Storing scanned documents

Correct Answer: 2. Providing a prediction immediately after each application request

Explanation:

An online machine-learning endpoint is appropriate when an application needs interactive predictions with low latency. For example, a web application could send information about a transaction to a deployed model and receive a prediction immediately so that the application can continue processing the request. Online endpoints are therefore suited to scenarios involving frequent individual requests and near-real-time responses. Batch endpoints are more appropriate when large collections of records can be processed together without requiring an immediate response for every record. Archiving files and storing documents are separate storage requirements and do not determine the need for online model inference.

Question 236. Which scenario is most suitable for batch inference?

1. Processing a large dataset periodically to generate predictions
2. Returning a prediction instantly for every website click
3. Providing live speech transcription
4. Translating a single sentence interactively

Correct Answer: 1. Processing a large dataset periodically to generate predictions

Explanation:

Batch inference is designed for processing large collections of records together, especially when immediate results for each individual request are not required. An organization might use batch inference to generate predictions for thousands or millions of records on a scheduled basis. For example, a company could process its customer database overnight and generate predictions for the following day’s activities. Online inference is better suited to interactive requests that require immediate responses. Live transcription and interactive translation involve different AI capabilities and response patterns. Therefore, periodic processing of a large dataset is a suitable scenario for batch inference.

Question 237. Which AI capability identifies whether a piece of text contains a person, organization, or location?

1. Image classification
2. Named entity recognition
3. Speech synthesis
4. Image segmentation

Correct Answer: 2. Named entity recognition

Explanation:

Named entity recognition identifies and categorizes meaningful entities found within text. Depending on the supported capabilities, entities can include people, organizations, locations, dates, quantities, and other categories. For example, in the sentence “Sara works at Contoso in London,” an NER system could identify Sara as a person, Contoso as an organization, and London as a location. This capability is useful when applications need to transform unstructured text into structured information. Image classification and segmentation analyze visual content, while speech synthesis generates spoken audio. Therefore, named entity recognition is the appropriate capability for identifying entities in written language.

Question 238. Which Azure AI capability converts text into natural-sounding spoken audio?

1. Speech-to-text
2. Text-to-speech
3. OCR
4. Semantic search

Correct Answer: 2. Text-to-speech

Explanation:

Text-to-speech converts written text into synthesized spoken audio. It can be used in virtual assistants, accessibility applications, automated announcements, educational systems, and other scenarios where information needs to be presented through speech. The application provides text, and the speech service generates audio using an appropriate supported voice and language. Speech-to-text performs the reverse process by converting spoken language into text. OCR extracts text from images, while semantic search retrieves information based on meaning. Therefore, text-to-speech is the appropriate capability when an AI application needs to turn written content into natural-sounding speech.

Question 239. Why should AI models be monitored after deployment?

1. Because real-world data and system performance can change
2. Because training becomes unnecessary
3. Because monitoring automatically guarantees fairness
4. Because all models expire after one day

Correct Answer: 1. Because real-world data and system performance can change

Explanation:

AI models operate in environments that can change after deployment. User behavior, business conditions, input data, and relationships between variables may evolve, potentially reducing model performance. Monitoring helps organizations identify these changes and determine whether further investigation, retraining, or other corrective action is needed. It can also help detect unusual behavior and provide information about the ongoing operation of the system. Monitoring does not guarantee fairness by itself and does not eliminate the need for training or evaluation. Therefore, models should be monitored because their real-world operating conditions and performance can change over time.

Question 240. What is a key reason to keep training, validation, and test datasets separate?

1. To ensure the model can be evaluated on data not used for training decisions
2. To make every model perfectly accurate
3. To eliminate all missing data automatically
4. To prevent users from accessing the application

Correct Answer: 1. To ensure the model can be evaluated on data not used for training decisions

Explanation:

Keeping training, validation, and test datasets separate helps provide a more reliable assessment of model performance. The training dataset is used to fit the model, while validation data can support development and tuning decisions. A test dataset is generally reserved for final evaluation using data that was not used to make those development decisions. If the same data is repeatedly used for training and evaluation, performance measurements may become overly optimistic. Separating datasets therefore helps assess how well the model generalizes to unseen examples. It does not guarantee perfect accuracy or automatically remove missing data.