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

 

Question 181. Which machine learning approach learns from data that contains known labels?

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

Correct Answer: 3. Supervised learning

Explanation:

Supervised learning uses a dataset in which the expected outcome, or label, is already known for each training example. The machine-learning algorithm learns relationships between the input features and the corresponding labels so it can make predictions for new data. Common supervised learning tasks include classification and regression. For example, a model can learn from historical house information where the sale price is known and then predict prices for new houses. Unsupervised learning does not require known labels, while clustering is an unsupervised technique. Reinforcement learning instead uses feedback from actions and outcomes. Therefore, supervised learning is the appropriate approach when training data contains known target labels.

Question 182. Which machine learning technique groups similar data points without predefined labels?

1. Regression
2. Classification
3. Supervised learning
4. Clustering

Correct Answer: 4. Clustering

Explanation:

Clustering is an unsupervised machine-learning technique that groups data points according to similarities in their characteristics. The training data does not need predefined labels describing the correct group for each record. For example, a business could use clustering to group customers according to purchasing behavior without first defining customer categories. The algorithm identifies patterns and creates groups based on the available features. Classification requires labeled examples and predicts predefined categories, while regression predicts numerical values. Therefore, clustering is useful when an organization wants to discover natural groupings in data rather than predict an already-known label.

Question 183. Which machine-learning task predicts a continuous numerical value?

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

Correct Answer: 1. Regression

Explanation:

Regression is a supervised machine-learning task used to predict continuous numerical values. Examples include predicting house prices, estimating monthly sales, forecasting energy consumption, or predicting delivery times. During training, the model learns relationships between input features and known numerical target values. After training, it can estimate the target value for previously unseen data. Classification is different because it predicts categories such as approved or rejected, while clustering groups similar observations without predefined labels. Entity recognition is a natural-language processing task that identifies meaningful entities in text. Therefore, regression is the appropriate machine-learning technique when the required output is a continuous number.

Question 184. Which metric measures the proportion of predictions that are correct overall?

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

Correct Answer: 2. Accuracy

Explanation:

Accuracy measures the proportion of predictions that a classification model gets correct across the evaluated dataset. It is calculated by dividing the number of correct predictions by the total number of predictions. For example, if a model correctly classifies 90 out of 100 records, its accuracy is 90 percent. Accuracy can be useful when the classes are reasonably balanced and the costs of different types of errors are similar. Precision focuses on the correctness of positive predictions, while recall measures how many actual positive cases were identified. Mean absolute error is generally used for numerical prediction tasks rather than classification accuracy.

Question 185. Which metric is especially useful when false positives are costly?

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

Correct Answer: 3. Precision

Explanation:

Precision measures how many of the cases predicted as positive were actually positive. It is particularly important when false positives create significant costs or undesirable consequences. For example, if an AI system flags transactions as potentially fraudulent, unnecessarily flagging legitimate transactions can create extra investigation work or inconvenience customers. A high-precision model reduces the proportion of incorrect positive predictions among all positive predictions. Recall has a different emphasis because it measures how many actual positive cases were successfully identified. Accuracy considers overall correctness, while mean squared error is generally associated with regression. The appropriate metric therefore depends on which types of errors matter most in the application.

Question 186. What does recall measure in a classification model?

1. The proportion of actual positive cases correctly identified
2. The average numerical prediction error
3. The number of training features
4. The percentage of storage used

Correct Answer: 1. The proportion of actual positive cases correctly identified

Explanation:

Recall measures the ability of a classification model to identify actual positive cases. It is calculated by comparing true positive predictions with all actual positive cases, including those that the model incorrectly classified as negative. Recall is particularly important when missing a positive case can have serious consequences. For example, in a system designed to detect a particular condition or security threat, failing to identify an actual positive case may be more problematic than generating some additional false alarms. Precision instead focuses on the correctness of positive predictions. Training features and storage usage are unrelated to the mathematical meaning of recall.

Question 187. Which metric is commonly used to measure the average absolute error of regression predictions?

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

Correct Answer: 4. Mean absolute error

Explanation:

Mean absolute error, commonly called MAE, measures the average absolute difference between predicted numerical values and their actual values. It is commonly used to evaluate regression models. For example, if a model predicts product demand and the predictions differ from actual demand by various amounts, MAE summarizes the average size of those errors without considering their direction. A lower MAE generally indicates that predictions are closer to the actual values. Accuracy, precision, and recall are commonly associated with classification tasks and do not directly measure average numerical prediction error. MAE is therefore appropriate when evaluating the typical magnitude of errors in continuous-value predictions.

Question 188. Why is a validation dataset used during model development?

1. To permanently store production predictions
2. To evaluate and tune a model during development
3. To replace the training dataset
4. To create user accounts

Correct Answer: 2. To evaluate and tune a model during development

Explanation:

A validation dataset is used during model development to evaluate how well a model performs on data that was not used directly to fit its parameters. Developers can use validation results to compare different model configurations, select hyperparameters, and identify potential overfitting. A separate test dataset is commonly reserved for a final evaluation after development decisions have been made. The validation dataset does not permanently store production predictions and does not replace the training data. Its purpose is to provide useful feedback during the development process so that the model configuration can be improved before final testing and deployment.

Question 189. What is overfitting in machine learning?

1. When a model learns the training data too closely and performs poorly on new data
2. When a model has no training data
3. When a database becomes unavailable
4. When a network connection is interrupted

Correct Answer: 1. When a model learns the training data too closely and performs poorly on new data

Explanation:

Overfitting occurs when a machine-learning model becomes too closely adapted to its training data, including patterns or noise that do not generalize to new examples. Such a model may show strong performance on the training dataset but significantly weaker performance on validation or test data. Developers can address overfitting through approaches such as using more representative training data, simplifying the model, applying regularization, or adjusting training procedures. Overfitting is a model-development issue rather than a database or network problem. The key characteristic is a gap between strong training performance and weaker performance on previously unseen data.

Question 190. What is underfitting?

1. A model that perfectly memorizes every training example
2. A model that has excessive storage capacity
3. A model that is too simple to capture important patterns
4. A model that only runs online

Correct Answer: 3. A model that is too simple to capture important patterns

Explanation:

Underfitting occurs when a machine-learning model is not sufficiently capable of representing the important patterns in the training data. The model may perform poorly not only on new data but also on the training dataset itself. This can happen when the model is too simple, important features are missing, or the training process is insufficient. Underfitting differs from overfitting, where the model learns the training data excessively and fails to generalize. Developers may address underfitting by improving features, selecting a more appropriate model, or adjusting training parameters. Therefore, a model that is too simple to capture meaningful patterns is an example of underfitting.

Question 191. Which Azure Machine Learning resource provides managed compute for training machine-learning models?

1. Compute cluster
2. Azure DNS zone
3. Azure Firewall
4. Content Safety policy

Correct Answer: 1. Compute cluster

Explanation:

An Azure Machine Learning compute cluster provides managed compute resources that can be used for machine-learning workloads such as model training. Compute clusters can scale according to workload requirements, helping organizations avoid manually managing individual machines for every training task. They are useful when training jobs require dedicated or scalable processing resources. Azure DNS zones are used for domain name resolution, Azure Firewall provides network security, and Content Safety policies address content moderation. These services have different purposes. Therefore, when an AI developer needs managed compute for machine-learning training workloads, an Azure Machine Learning compute cluster is an appropriate resource.

Question 192. What is a feature in a supervised machine-learning dataset?

1. The final model deployment
2. An input variable used by the model
3. A network security rule
4. The model’s endpoint URL

Correct Answer: 2. An input variable used by the model

Explanation:

A feature is an input variable that provides information the machine-learning model can use to make a prediction. For example, when predicting house prices, features might include the number of bedrooms, property size, location, and age of the building. The model learns relationships between these input features and the target label during training. The label represents the expected outcome that the model is attempting to predict. Features therefore form an important part of the input data used by supervised learning algorithms. Deployment URLs and security rules are infrastructure concepts and are not considered machine-learning features.

Question 193. Which practice can help reduce bias in an AI model?

1. Removing all evaluation data
2. Ignoring differences between groups
3. Evaluating performance across relevant demographic groups
4. Using fewer test cases regardless of quality

Correct Answer: 3. Evaluating performance across relevant demographic groups

Explanation:

Evaluating an AI model across relevant demographic or user groups can help identify differences in performance and potential sources of unfair outcomes. A model may perform well overall while producing significantly different results for particular groups. Reviewing evaluation results across groups can reveal such disparities and encourage developers to investigate data quality, representation, features, or model behavior. Bias reduction may involve several additional steps, including improving training data and reviewing the model’s intended use. Simply ignoring group differences or removing evaluation data makes potential problems harder to detect. Therefore, group-based evaluation is an important practice when assessing fairness.

Question 194. Which principle requires an organization to be able to explain important information about an AI system?

1. Transparency
2. Clustering
3. Regression
4. Encryption

Correct Answer: 1. Transparency

Explanation:

Transparency concerns communicating relevant information about an AI system so that users and stakeholders can understand important aspects of how it is developed and used. Depending on the application, this can include information about the system’s purpose, capabilities, limitations, data usage, or evaluation. Transparency can help people make informed decisions about when and how to rely on an AI system. It is distinct from fairness, which focuses on equitable treatment, and accountability, which concerns responsibility for the system and its outcomes. Clustering and regression are machine-learning techniques, while encryption is a security mechanism rather than a responsible AI principle.

Question 195. What is the primary purpose of model monitoring after deployment?

1. To permanently eliminate testing
2. To detect changes in performance or data
3. To create training labels automatically in every case
4. To replace the model with a database

Correct Answer: 2. To detect changes in performance or data

Explanation:

Model monitoring helps organizations observe deployed AI systems and identify changes that may affect their performance or reliability. After deployment, real-world data and user behavior can change, meaning a model that performed well during development may gradually become less effective. Monitoring can help identify changes in input distributions, prediction patterns, performance metrics, or other relevant signals. When problems are detected, an organization can investigate whether retraining, recalibration, or another update is appropriate. Monitoring does not eliminate the need for testing or automatically replace the model with a database. Its main purpose is to provide visibility into the ongoing behavior of the deployed AI system.

Question 196. Which inference method is appropriate for processing a large collection of records without an immediate response for each request?

1. Online inference
2. Manual inference
3. Batch inference
4. Interactive prompting only

Correct Answer: 3. Batch inference

Explanation:

Batch inference is designed for processing many records as a group rather than responding immediately to each individual request. It is useful when an organization needs to generate predictions for a large dataset on a scheduled or periodic basis. For example, a business could process thousands of customer records overnight and generate predictions for the next day’s operations. Online inference is better suited to interactive scenarios where an application expects a quick response for each request. Batch processing can therefore be more efficient when immediate individual responses are not required. The choice depends on workload characteristics, response-time requirements, and application architecture.

Question 197. Which statement best describes generalization in machine learning?

1. The ability of a model to perform well on previously unseen data
2. The process of deleting training data
3. The creation of a network subnet
4. The storage of application logs

Correct Answer: 1. The ability of a model to perform well on previously unseen data

Explanation:

Generalization refers to a machine-learning model’s ability to apply learned patterns successfully to new data that was not part of the training process. A model that generalizes well does not simply memorize its training examples; instead, it learns useful relationships that remain applicable to future observations. Good training, representative datasets, appropriate model complexity, and proper evaluation can all contribute to better generalization. Overfitting is a common problem because an overfit model may perform very well on training data but poorly on unseen examples. Network subnets and application logs are infrastructure concepts and are unrelated to the definition of model generalization.

Question 198. Which Azure capability can help evaluate generative AI responses against defined criteria?

1. Azure DNS
2. Azure AI evaluation capabilities
3. Azure Virtual Network
4. Azure Storage account keys

Correct Answer: 2. Azure AI evaluation capabilities

Explanation:

AI evaluation capabilities help developers assess the quality, relevance, safety, and other characteristics of generative AI outputs against defined criteria. Evaluation is important because a generative model can produce different responses for similar requests and may occasionally generate inaccurate or unsuitable content. Developers can establish evaluation datasets and metrics that reflect the application’s requirements, then use the results to compare prompts, models, or application configurations. This provides a more systematic approach than relying only on individual manual observations. Azure DNS, virtual networking, and storage account keys serve infrastructure or security functions and are not designed to evaluate the quality of generated responses.

Question 199. Which action can help protect an AI application from unauthorized access to its resources?

1. Disabling all authentication
2. Publishing secrets in prompts
3. Applying appropriate identity and access controls
4. Sharing administrator credentials

Correct Answer: 3. Applying appropriate identity and access controls

Explanation:

Identity and access controls help ensure that only authorized users, applications, or services can access AI resources and related data. Proper access management can include authentication, authorization, role-based permissions, managed identities, and the principle of least privilege. These controls are especially important when an AI application can access private documents, databases, models, or other organizational resources. Disabling authentication or sharing administrator credentials increases security risk. Secrets should also be protected rather than placed directly into prompts or application content. Therefore, applying appropriate identity and access controls is an important security practice for AI solutions.

Question 200. Why should an AI solution be tested with data that represents its expected real-world users and scenarios?

1. To guarantee that every prediction is correct
2. To eliminate the need for monitoring
3. To make the model independent of its training data
4. To evaluate whether the system performs appropriately in its intended context

Correct Answer: 4. To evaluate whether the system performs appropriately in its intended context

Explanation:

Testing an AI solution with representative data helps determine whether the system performs appropriately for the users, conditions, and scenarios it is expected to encounter after deployment. A dataset that does not reflect real-world usage may hide important performance problems or differences between groups. Representative evaluation can help reveal issues related to accuracy, fairness, reliability, and generalization before the system is widely used. Testing cannot guarantee that every prediction will be correct, and it does not eliminate the need for monitoring after deployment. Instead, it provides evidence about how well the AI solution is likely to behave within its intended application context.