View Full Microsoft AI-900 Exam Dumps and Practice Test Dumps.

 

Question 381

A robot learns to choose actions by receiving rewards when it performs desirable actions. Which machine learning approach is being used?

1. Supervised learning
2. Unsupervised learning
3. Regression
4. Reinforcement learning

Correct Answer: 4

Explanation

Reinforcement learning trains an agent to make decisions by interacting with an environment and receiving feedback in the form of rewards or penalties. The agent gradually learns which actions are more useful for achieving a desired objective. Supervised learning relies on labeled training examples, while unsupervised learning discovers patterns without labeled outcomes. Regression predicts numerical values rather than learning through reward-based interaction. Therefore, when a robot improves its behavior based on rewards received after taking actions, the machine learning approach is reinforcement learning.

Question 382

A retailer has customer purchase records but no predefined categories and wants to discover groups of customers with similar buying behavior. Which approach is appropriate?

1. Unsupervised learning
2. Supervised learning
3. Speech recognition
4. Regression

Correct Answer: 1

Explanation

Unsupervised learning is appropriate when a dataset does not contain predefined labels and the goal is to discover meaningful patterns or groups within the data. Clustering is a common unsupervised learning technique that can group customers according to similarities in their purchasing behavior. Supervised learning requires labeled examples, while speech recognition processes spoken language. Regression predicts numerical values based on input features. Therefore, because the retailer wants to discover natural customer groups without predefined categories, unsupervised learning is the appropriate approach.

Question 383

A bank wants to predict whether a loan applicant will default based on income, credit history, and existing debt. The result should be either “default” or “no default.” Which machine learning task should be selected?

1. Regression
2. Classification
3. Clustering
4. Anomaly detection

Correct Answer: 2

Explanation

Classification is used when a machine learning model predicts a category or class from input features. In this scenario, the desired outcome consists of two categories: default and no default. The applicant’s income, credit history, and existing debt can be used as features for making the prediction. Regression is intended for numerical outputs, clustering discovers groups without predefined labels, and anomaly detection identifies unusual observations. Therefore, because the model must select between two predefined categories, this scenario represents a classification task.

Question 384

A machine learning model predicts the monthly electricity consumption of a building as a numerical value. Which task is being performed?

1. Classification
2. Clustering
3. Regression
4. Entity recognition

Correct Answer: 3

Explanation

Regression is a machine learning task used to predict continuous numerical values. Monthly electricity consumption can be represented by a numerical quantity, such as the number of kilowatt-hours consumed. A regression model can learn relationships between input features, such as building size, occupancy, weather, and historical usage, and the target consumption value. Classification produces categories, clustering identifies groups, and entity recognition extracts named entities from text. Therefore, when the objective is to predict a numerical electricity-consumption value, regression is the suitable machine learning task.

Question 385

A data scientist wants to measure how far a regression model’s predictions are from the actual numerical values on average. Which metric can be used?

1. Accuracy
2. Recall
3. F1 score
4. Mean absolute error

Correct Answer: 4

Explanation

Mean absolute error, or MAE, measures the average absolute difference between predicted values and actual values. It is commonly used to evaluate regression models because it expresses the average prediction error in the same general units as the target variable. Accuracy, recall, and F1 score are commonly associated with classification evaluation rather than continuous numerical prediction. A lower MAE generally indicates that predictions are closer to the actual values. Therefore, mean absolute error is an appropriate metric for measuring average prediction error in this regression scenario.

Question 386

A model predicts whether transactions are fraudulent. The team is especially concerned about incorrectly classifying legitimate transactions as fraudulent because this could inconvenience customers. Which type of error should the team monitor closely?

1. False positive
2. False negative
3. True positive
4. True negative

Correct Answer: 1

Explanation

A false positive occurs when a model predicts the positive class even though the actual case belongs to the negative class. In fraud detection, if fraudulent transactions are considered positive and legitimate transactions are negative, a false positive means a legitimate transaction is incorrectly flagged as fraudulent. This can inconvenience customers or trigger unnecessary investigations. A false negative is the opposite type of classification error. Therefore, when the primary concern is incorrectly flagging legitimate transactions as fraud, the team should closely monitor false positives.

Question 387

A search system represents documents and user queries as numerical vectors so that it can compare their semantic similarity. What are these numerical representations commonly called?

1. Labels
2. Embeddings
3. Confusion matrices
4. Hyperparameters

Correct Answer: 2

Explanation

Embeddings are numerical representations of data that capture meaningful relationships or semantic characteristics. In search scenarios, text such as documents and user queries can be converted into vectors, allowing the system to compare their semantic similarity. This can help retrieve relevant content even when the wording differs between the query and document. Labels represent target values, confusion matrices summarize classification results, and hyperparameters are configuration values used during model development. Therefore, numerical vector representations used for semantic comparison are commonly called embeddings.

Question 388

A company wants a search application to retrieve information based on the meaning of a question instead of requiring the same keywords to appear in a document. Which capability is most relevant?

1. OCR
2. Speech synthesis
3. Semantic search
4. Image classification

Correct Answer: 3

Explanation

Semantic search focuses on the meaning and intent behind a query rather than relying only on exact keyword matches. This allows a search system to identify content that is conceptually related to the user’s question, even when different words are used in the stored documents. OCR extracts text from images, speech synthesis converts text into spoken audio, and image classification categorizes images. Therefore, when a company wants search results based on the meaning of a user’s question, semantic search is the relevant capability.

Question 389

An AI assistant generates an answer that sounds convincing but contains information that cannot be supported by the organization’s source documents. What issue has occurred?

1. Overfitting
2. Data normalization
3. Hallucination
4. Clustering

Correct Answer: 3

Explanation

A hallucination occurs when a generative AI system produces information that may sound plausible but is inaccurate, unsupported, or not grounded in reliable source information. This can be particularly problematic in business applications where answers should be based on trusted organizational content. Overfitting is a machine learning issue involving excessive adaptation to training data, while data normalization concerns data preparation. Clustering groups similar data points. Therefore, when an AI assistant produces convincing but unsupported information, the issue is a generative AI hallucination.

Question 390

A company wants a generative AI application to answer questions using information retrieved from its internal knowledge base. Which architecture can provide the retrieved information as context for generation?

1. Retrieval-augmented generation
2. Image classification
3. Regression
4. Optical character recognition

Correct Answer: 1

Explanation

Retrieval-augmented generation, or RAG, combines information retrieval with generative AI. When a user asks a question, relevant information can first be retrieved from an approved knowledge source and then supplied to the generative model as context for producing the response. This approach can help ground generated answers in organizational information and reduce dependence on information contained only within the model. Image classification categorizes images, regression predicts numerical values, and OCR extracts text from images. Therefore, RAG is suitable for this knowledge-based generative AI scenario.

Question 391

An organization wants to extract invoice numbers, vendor names, dates, and total amounts from many invoices that follow different layouts. Which Azure AI capability is most appropriate when a custom document model is required?

1. Azure AI Speech
2. Azure AI Vision
3. Azure AI Language
4. Azure AI Document Intelligence

Correct Answer: 4

Explanation

Azure AI Document Intelligence is designed to analyze documents and extract structured information from forms, invoices, receipts, and other document types. When documents have organization-specific layouts or extraction requirements, a custom model can be created to recognize and extract the required fields. Azure AI Speech handles spoken-language workloads, Azure AI Vision focuses on image analysis, and Azure AI Language processes text and natural language. Therefore, Document Intelligence is the appropriate Azure AI capability for extracting structured invoice information using a custom document model.

Question 392

A company wants to automatically assign incoming customer emails to categories such as Billing, Technical Support, and Sales. Which language workload is most suitable?

1. Text classification
2. Image captioning
3. Speech synthesis
4. Object detection

Correct Answer: 1

Explanation

Text classification assigns written content to predefined categories based on its meaning or characteristics. In this scenario, incoming customer emails can be classified into categories such as Billing, Technical Support, and Sales. A trained classification model can learn patterns in existing examples and use them to categorize new messages. Image captioning describes visual content, speech synthesis converts text to spoken audio, and object detection identifies objects in images. Therefore, because the emails must be assigned to predefined categories, text classification is the suitable language workload.

Question 393

A model performs very well on its training dataset but performs poorly when given new data. Which machine learning problem is most likely present?

1. Underfitting
2. Overfitting
3. Clustering
4. Feature extraction

Correct Answer: 2

Explanation

Overfitting occurs when a model learns the training data too closely, including patterns or noise that do not generalize well to new examples. As a result, the model may show excellent performance on training data while producing significantly weaker results on unseen data. Underfitting occurs when a model is too simple to capture important patterns in the data. Clustering is an unsupervised learning technique, and feature extraction involves preparing useful input information. Therefore, strong training performance combined with poor performance on new data is a common sign of overfitting.

Question 394

A model performs poorly on both its training data and its test data because it is too simple to capture important patterns. Which problem does this describe?

1. Data leakage
2. Overfitting
3. Underfitting
4. Vector search

Correct Answer: 3

Explanation

Underfitting occurs when a machine learning model is too simple or insufficiently trained to capture important relationships in the data. Because the model has not learned the underlying patterns effectively, it can perform poorly on both the training dataset and unseen test data. Overfitting produces a different pattern, where training performance may be strong while generalization is poor. Data leakage involves inappropriate information entering the training process, and vector search is a retrieval technique. Therefore, poor performance on both training and test data due to excessive simplicity indicates underfitting.

Question 395

A machine learning team discovers that information from the test dataset was accidentally included during model training. Why is this a problem?

1. It can cause data leakage and make evaluation unreliable
2. It always improves fairness
3. It converts regression into clustering
4. It removes the need for validation

Correct Answer: 1

Explanation

Including test data during training can cause data leakage because information intended to remain unseen becomes available to the model during development. This can make evaluation results appear better than the model’s true ability to generalize to new data. A properly separated test dataset should be reserved for final evaluation so that it provides an independent estimate of performance. Data leakage does not automatically improve fairness, change the type of machine learning task, or eliminate the need for validation. Therefore, test-data leakage can make model evaluation unreliable.

Question 396

An AI application processes personal customer information. The development team wants to ensure that only authorized users and services can access the data. Which measure directly supports this goal?

1. Increasing image resolution
2. Applying access controls
3. Increasing model temperature
4. Adding more output tokens

Correct Answer: 2

Explanation

Access controls help restrict data and system resources to authorized users, applications, or services. In AI solutions that process personal customer information, appropriate identity and access management can reduce the risk of unauthorized access to sensitive information. Image resolution affects visual quality, model temperature influences generative output behavior, and output-token limits control the amount of generated text. None of those directly controls who can access protected data. Therefore, applying appropriate access controls is a direct measure for protecting customer information.

Question 397

An AI team wants to ensure that a model provides comparable performance for different demographic groups. What responsible AI principle should receive particular attention?

1. Fairness
2. Compression
3. Latency
4. Indexing

Correct Answer: 1

Explanation

Fairness focuses on ensuring that AI systems do not produce unjustified differences in treatment or outcomes for different groups. A team evaluating model performance across demographic groups can compare relevant metrics and investigate whether systematic disparities exist. Compression is concerned with reducing model size, latency concerns response speed, and indexing is associated with organizing information for retrieval. These may be important technical considerations but do not directly address differences in model outcomes between groups. Therefore, fairness is the responsible AI principle most relevant to this requirement.

Question 398

A developer wants a generative AI system to follow specific instructions about its role, behavior, and response style throughout a conversation. Which prompt component is most appropriate for defining these high-level instructions?

1. Search index
2. System message
3. Test label
4. Feature vector

Correct Answer: 2

Explanation

A system message provides high-level instructions that guide how a generative AI model should behave. It can define the assistant’s role, establish behavioral rules, specify response requirements, and provide other instructions that should influence the conversation. A search index organizes information for retrieval, a test label represents an expected outcome in supervised learning, and a feature vector represents numerical input information. Therefore, when the developer needs to establish persistent behavioral guidance for a generative AI application, a system message is an appropriate prompt component.

Question 399

A company evaluates an AI solution using only a small set of unusually easy examples and then concludes that the model is highly accurate. What should the team do to obtain a more meaningful evaluation?

1. Remove all test data
2. Use representative evaluation data
3. Train only on the easiest examples
4. Ignore performance differences

Correct Answer: 2

Explanation

Representative evaluation data should reflect the types of inputs and conditions the AI system is expected to encounter in real-world use. Evaluating a model only on unusually easy examples can produce misleadingly high performance and fail to reveal important weaknesses. A suitable evaluation dataset should contain relevant variation and realistic cases, including challenging examples where appropriate. Removing test data or training only on easy examples does not provide reliable evidence of generalization. Therefore, using representative evaluation data is important for obtaining a meaningful assessment of an AI solution.

Question 400

A company is assessing a new AI solution before deploying it. The team checks performance, privacy, fairness, security, and the system’s ability to operate reliably under expected conditions. Why is this broader evaluation important?

1. AI quality depends only on model accuracy
2. Responsible AI applies only after deployment
3. AI solutions should be assessed across multiple relevant requirements
4. Technical metrics are unnecessary

Correct Answer: 3

Explanation

Evaluating an AI solution involves more than measuring predictive accuracy. A reliable assessment can consider performance, fairness, privacy, security, transparency, inclusiveness, and reliability and safety, depending on the scenario. These considerations help identify risks that may not be visible through a single technical metric. Responsible AI principles should be considered throughout the solution lifecycle rather than only after deployment. Therefore, a broader evaluation is important because AI systems should be assessed against multiple relevant technical, operational, and responsible AI requirements before they are used in real-world scenarios.