AI-300 Practice Questions & Answers

Loading your practice set…

In designing an MLOps architecture for a machine learning project, you need to integrate a feature store to manage and provide features for model training. Which of the following best describes the role of a feature store in this context?

Data Lake for storing raw data

Centralized repository for managing and providing features (Correct)

Data Warehouse for analytics

Data Catalog for metadata discovery

Explanation: The correct answer is B. A feature store centralizes the storage and management of features used in machine learning models, ensuring that the same features are used across different models and teams. This helps in maintaining consistency and reproducibility. Option A is incorrect because it refers to a data lake, which is primarily for storing raw data. Option C is incorrect as it pertains to a data warehouse, which is used for analytics rather than feature storage. Option D is incorrect because it refers to a data catalog, which is used for metadata discovery and governance, not feature storage.

When setting up a CI/CD pipeline for a machine learning project, you want to ensure that the environment used for training and testing is consistent and reproducible. Which of the following tools or services would best help achieve this goal?

Azure DevOps for version control and CI/CD

Azure Machine Learning Workspaces for reproducible environments (Correct)

Kubernetes for container orchestration

Docker images for containerization

Explanation: The correct answer is B. Azure Machine Learning Workspaces provide a managed environment for building, training, and deploying machine learning models, including reproducible environments. Option A is incorrect because Azure DevOps is primarily for version control and CI/CD pipelines, not for managing machine learning environments. Option C is incorrect as it refers to Kubernetes, which is used for container orchestration but not specifically for reproducible ML environments. Option D is incorrect because Docker images can be used for containerization but do not provide the full suite of tools for reproducible ML environments.

You are tasked with automating the model training process using GitHub Actions for Azure Machine Learning. You want to ensure that the model is validated before it is deployed to production. Which of the following steps should you include in your GitHub Actions workflow to achieve this?

Train the model using Azure ML SDK and then deploy it directly to production.

Use Azure ML SDK to train the model, validate it using a test dataset, and then deploy it to production if the validation passes. (Correct)

Deploy the model to production and then validate it using a test dataset.

Use Azure ML SDK to train the model and then manually validate it outside of the workflow.

Explanation: The correct answer is B because it includes both the training and validation steps within the same workflow. Option A only trains the model without validation, which does not meet the requirement. Option C only validates the model without training, which is not sufficient. Option D does not mention any validation or training steps, making it incorrect.

During the deployment of an AI model, you decide to use canary deployment to minimize risk. Which of the following best describes the canary deployment process in the context of Azure Machine Learning and GitHub Actions?

Deploy the entire model to all users immediately after training.

Deploy a small portion of the traffic to the new model version to test its performance before fully rolling out. (Correct)

Manually test the model on a subset of data before deploying it to production.

Deploy the model to a separate environment for testing but not to production.

Explanation: The correct answer is B because it accurately describes the canary deployment process where a small percentage of traffic is directed to the new model version to test its performance before full deployment. Option A is incorrect as it suggests deploying the entire model at once, which contradicts the purpose of canary deployment. Option C is incorrect as it implies that canary deployment is not automated, which is not true. Option D is incorrect as it suggests that canary deployment is not used for testing, which is a key aspect of canary deployment.

In an Azure Machine Learning deployment, you notice that the performance of your model has degraded over time. You want to monitor for data drift to understand if changes in the input data have caused this degradation. Which Azure service should you use to set up a data drift monitor and how should you configure it?

Use Azure Monitor with custom logs

Set up an Azure Machine Learning Data Drift Monitor with a specified baseline dataset and drift check frequency (Correct)

Deploy a custom script using Azure Functions to periodically check for drift

Configure an Azure DevOps pipeline to run model validation tests

Explanation: The correct answer is to use the Azure Machine Learning Data Drift Monitor. It allows you to specify the baseline dataset and the frequency at which the drift should be checked. The other options are incorrect because they do not provide the specific functionality required for monitoring data drift.

Your organization wants to ensure that the features used in their machine learning models are contributing meaningfully to the predictions. They are particularly interested in understanding how each feature impacts the model's output. Which tool or method would you recommend to achieve this, and what is the key benefit of using this approach?

Implement a confusion matrix analysis

Use SHAP (SHapley Additive exPlanations) for feature attribution (Correct)

Perform a ROC curve analysis

Apply cross-validation techniques

Explanation: The correct answer is to use feature attribution methods such as SHAP (SHapley Additive exPlanations). This method provides insights into how each feature contributes to the model's output, which helps in understanding the model's behavior and ensuring that the features are meaningful. The other options are incorrect because they do not provide the specific method for feature attribution.