View Full PMI CPMAI Exam Dumps and Practice Test Dumps.

 

Question 381

During the early discovery stage of an AI project, what should the project team establish first?

1. The production deployment pipeline
2. The business problem and desired outcome
3. The final model hyperparameters
4. The model retirement schedule

Correct Answer: 2

Explanation

The discovery stage should begin by clearly understanding the business problem and the outcome the organization expects to achieve. This provides the foundation for determining whether AI is appropriate and what type of solution may be required. Starting with technical details such as hyperparameters or deployment pipelines can lead to building a technically sophisticated system that does not address the actual business need. A clearly defined problem also helps establish scope, stakeholders, measurable objectives, data requirements, and success criteria. Once the desired outcome is understood, the team can evaluate feasibility and determine the most suitable AI approach.

Question 382

Which practice is most useful when evaluating a potential data source for an AI project?

1. Checking its relevance, quality, availability, and reliability
2. Selecting the source with the largest file size
3. Choosing only data that requires no documentation
4. Using the newest source regardless of its relevance

Correct Answer: 1

Explanation

A potential data source should be evaluated systematically before it becomes part of an AI project. Important considerations include whether the data is relevant to the business problem, sufficiently accurate and complete, available when needed, and reliable over time. The source should also be assessed for ownership, licensing, privacy requirements, accessibility, and update frequency. Large datasets are not automatically better, and newer data may still be irrelevant or unreliable. A structured source evaluation helps prevent downstream problems involving data quality, availability, compliance, and model performance. It also supports better decisions about whether additional sources or data acquisition methods are necessary.

Question 383

What is a key responsibility of a data steward in an AI initiative?

1. Selecting the final machine learning algorithm
2. Writing all model inference code
3. Managing data quality, definitions, access, and governance practices
4. Approving every production deployment independently

Correct Answer: 3

Explanation

A data steward helps ensure that organizational data is properly managed and remains suitable for its intended uses. Responsibilities can include supporting data quality, maintaining definitions and metadata, clarifying ownership, monitoring access practices, and helping enforce data governance policies. The data steward does not normally own the machine learning algorithm or independently approve every deployment. Clear stewardship is especially important in AI projects because poor data quality, unclear definitions, inappropriate access, or missing lineage can negatively affect model outcomes and compliance. Effective stewardship creates greater confidence that project teams understand where data originated, what it represents, and how it may appropriately be used.

Question 384

Why is inter-annotator agreement useful when creating labeled training data?

1. It measures how quickly annotators complete their work
2. It determines the model’s inference latency
3. It automatically selects the best algorithm
4. It helps assess consistency between human labelers

Correct Answer: 4

Explanation

Inter-annotator agreement evaluates how consistently different human annotators assign labels to the same or similar examples. High agreement can indicate that labeling guidelines are sufficiently clear and that the task is interpreted consistently. Low agreement may reveal ambiguous definitions, inadequate instructions, difficult examples, or subjective judgment. Addressing these issues before training can improve label quality and reduce noise in the dataset. Depending on the task, different statistical measures can be used to quantify agreement. The purpose is not to measure annotation speed or model performance directly, but to understand the consistency and reliability of the human-generated labels used during AI development.

Question 385

What is an important consideration when using synthetic data for AI model development?

1. Whether it adequately represents relevant real-world characteristics
2. Whether it completely eliminates the need for real data
3. Whether it always produces higher accuracy
4. Whether it removes every possible source of bias

Correct Answer: 1

Explanation

Synthetic data can help address limited data availability, privacy concerns, or expensive data collection, but its usefulness depends on how realistically it represents the characteristics of the intended environment. Teams should evaluate whether synthetic examples capture important distributions, relationships, edge cases, and variations found in real-world data. Poorly generated synthetic data can introduce unrealistic patterns or reinforce existing assumptions. It should therefore be validated against appropriate real-world references whenever possible. Synthetic data does not automatically eliminate bias, guarantee higher model accuracy, or completely replace authentic data. Its suitability should be determined according to the specific AI use case and validation requirements.

Question 386

Which technique is specifically designed to reduce the impact of very large differences in numerical feature scales?

1. Clustering
2. Feature scaling
3. Label encoding
4. Data deduplication

Correct Answer: 2

Explanation

Feature scaling transforms numerical variables so that differences in their original units or magnitudes do not disproportionately influence an algorithm. Common approaches include standardization and normalization. This can be particularly important for algorithms that rely on distances, gradients, or numerical optimization. For example, a feature measured in thousands could otherwise dominate another feature measured between zero and one. Scaling does not change the underlying business meaning of the feature; it changes its numerical representation for modeling purposes. The appropriate technique depends on the algorithm and data distribution. Teams should document preprocessing steps so that the same transformations are consistently applied during validation and production inference.

Question 387

A project team wants to automatically search many possible hyperparameter combinations. Which approach is appropriate?

1. Manual data labeling
2. Feature deletion
3. Hyperparameter optimization
4. Model retirement

Correct Answer: 3

Explanation

Hyperparameter optimization systematically evaluates different configurations of model settings to identify combinations that perform well according to selected validation criteria. Techniques can include grid search, random search, and more advanced optimization methods. Hyperparameters are different from parameters learned directly from training data, and they can significantly affect model performance, training time, and generalization. A proper validation strategy is important so that the optimization process does not simply overfit a validation set. The project team should also record tested configurations and results to support reproducibility. Hyperparameter optimization is therefore a structured experimentation activity rather than a substitute for defining the business problem or evaluating data quality.

Question 388

Which metric is especially useful for evaluating the ability of a binary classifier to distinguish positive and negative classes across different thresholds?

1. Mean absolute error
2. Recall only
3. Accuracy only
4. ROC-AUC

Correct Answer: 4

Explanation

ROC-AUC summarizes a classifier’s ability to distinguish between positive and negative classes across different classification thresholds. The ROC curve compares the true-positive rate with the false-positive rate at different threshold settings, while the area under that curve provides an aggregate measure of discrimination. A higher ROC-AUC generally indicates stronger ranking or separation ability, although the metric should not be interpreted in isolation. In imbalanced datasets or situations where specific error types have greater business consequences, precision-recall analysis and threshold-specific metrics may provide additional insight. The appropriate evaluation approach should therefore reflect the project’s objectives, class distribution, and operational decision requirements.

Question 389

What is the primary purpose of regularization in machine learning?

1. To reduce overfitting by discouraging overly complex models
2. To guarantee perfect predictions on training data
3. To increase the number of missing values
4. To remove the need for validation data

Correct Answer: 1

Explanation

Regularization is used to reduce overfitting by discouraging models from becoming unnecessarily complex. Techniques such as L1 and L2 regularization add penalties related to model parameters during training. This encourages the model to focus on patterns that are more likely to generalize instead of memorizing noise in the training dataset. Regularization does not guarantee perfect performance and does not eliminate the need for validation or testing. Its effectiveness depends on choosing appropriate regularization settings and evaluating the resulting model on data that was not used for fitting. It is one of several techniques available for improving generalization and controlling model complexity.

Question 390

When should a classification threshold be adjusted?

1. Only when the training dataset contains images
2. When business consequences make a different precision-recall tradeoff appropriate
3. Only after the model has been retired
4. Whenever the model has zero training errors

Correct Answer: 2

Explanation

A classification threshold determines when a model’s predicted probability is converted into a particular class decision. The default threshold may not be appropriate when false positives and false negatives have different business consequences. For example, a project may prefer higher recall when missing a positive case is especially costly, even if that produces more false positives. Conversely, high precision may be more important when false alarms create substantial operational costs. Threshold selection should therefore be based on validation results, business requirements, risk tolerance, and operational constraints. Adjusting the threshold can change classification behavior without necessarily retraining the underlying model.

Question 391

Which approach can help evaluate whether an AI model behaves differently across relevant demographic or operational groups?

1. Increasing the training batch size
2. Removing all categorical variables
3. Comparing performance metrics across defined subgroups
4. Measuring only overall accuracy

Correct Answer: 3

Explanation

Comparing model performance across relevant subgroups can reveal differences that may be hidden by aggregate metrics. Teams can examine measures such as precision, recall, false-positive rates, false-negative rates, or other appropriate indicators for each defined group. The groups should be selected based on the use case, applicable requirements, and responsible AI considerations. A model can have acceptable overall performance while producing materially different outcomes for particular populations. Subgroup analysis does not automatically establish that a system is unfair, but it provides evidence that can guide further investigation. Results should be interpreted carefully, particularly when subgroup sample sizes are small or data quality differs between groups.

Question 392

What is the main purpose of a model registry in an AI lifecycle?

1. To permanently store raw user passwords
2. To replace all project documentation
3. To automatically define business objectives
4. To manage and track model versions and related metadata

Correct Answer: 4

Explanation

A model registry provides a centralized mechanism for managing model artifacts and associated metadata throughout the AI lifecycle. Depending on the platform, it may track versions, stages, ownership, evaluation results, deployment status, and related information. This helps teams identify which model is currently approved, which version is deployed, and which earlier versions are available for rollback or comparison. A registry supports governance, reproducibility, and operational control, but it does not replace broader documentation, data governance, or business requirements. Proper registry practices also make model changes easier to audit and communicate across development, validation, operations, and governance teams.

Question 393

Which deployment strategy releases a new AI model to a small portion of users or traffic before broader rollout?

1. Canary deployment
2. Full replacement deployment
3. Permanent offline evaluation
4. Data archival

Correct Answer: 1

Explanation

Canary deployment introduces a new model to a limited portion of production traffic before expanding its use. The team can monitor technical and business indicators such as latency, error rates, prediction quality, user behavior, and unexpected outcomes. If significant problems appear, traffic can be redirected to the previous model while the issue is investigated. This reduces the potential impact of an unsuccessful release compared with immediately deploying the new model to everyone. Canary deployment is particularly useful when production behavior may differ from testing environments. It should be supported by predefined monitoring thresholds, rollback procedures, and clear criteria for expanding or stopping the rollout.

Question 394

What is a major security concern specific to generative AI applications that accept user prompts?

1. Increasing storage capacity
2. Prompt injection
3. Reducing model size
4. Improving batch processing speed

Correct Answer: 2

Explanation

Prompt injection occurs when crafted input attempts to manipulate a generative AI system into ignoring intended instructions, revealing protected information, or performing actions outside its authorized purpose. This risk is particularly important when an AI application can access tools, private documents, databases, or external systems. Effective defenses can include separating trusted instructions from untrusted content, restricting tool permissions, validating outputs, limiting data access, monitoring suspicious inputs, and requiring human approval for sensitive actions. Prompt injection cannot necessarily be solved by relying on a single instruction in the system prompt. Security controls should instead be applied across the application architecture and AI lifecycle.

Question 395

What is the primary purpose of retrieval-augmented generation (RAG)?

1. To eliminate the need for any source documents
2. To replace all machine learning models with databases
3. To provide relevant retrieved information to support generated responses
4. To guarantee that every generated statement is correct

Correct Answer: 3

Explanation

Retrieval-augmented generation combines information retrieval with generative AI. Before generating a response, the system retrieves relevant content from an approved knowledge source and provides that content as context to the language model. This can help the model produce responses that are grounded in organizational or domain-specific information rather than relying only on information encoded during model training. RAG does not guarantee correctness because retrieval can return incomplete or irrelevant material and the model can still generate inaccurate statements. Effective implementations therefore require quality source data, retrieval evaluation, access controls, monitoring, and output validation appropriate to the application’s risk level.

Question 396

Which control is most appropriate for limiting an AI application’s access to only the data and systems required for its assigned function?

1. Least privilege
2. Unlimited administrator access
3. Shared credentials
4. Public data exposure

Correct Answer: 1

Explanation

The principle of least privilege limits users, services, and AI applications to only the permissions necessary to perform their authorized functions. This reduces the potential impact of compromised credentials, malicious inputs, software defects, or unintended model behavior. In an AI application, least privilege can involve restricting database access, limiting available tools, controlling document repositories, and separating read and write permissions. It should be combined with authentication, authorization, monitoring, and periodic access reviews. Giving an AI service broad administrator privileges increases the consequences of security failures. Least privilege is therefore an important architectural control for reducing unnecessary exposure throughout the AI system lifecycle.

Question 397

Which practice best supports reproducibility when machine learning experiments are performed by multiple team members?

1. Allowing each member to use undocumented settings
2. Recording datasets, code, parameters, and experiment results
3. Deleting unsuccessful experiments immediately
4. Changing environments without tracking versions

Correct Answer: 2

Explanation

Reproducibility requires enough information for another team member to understand and recreate an experiment. Useful records include dataset versions, preprocessing steps, source code, model configuration, hyperparameters, software dependencies, random seeds where relevant, evaluation results, and environment information. Experiment tracking systems can help organize these details and associate results with specific configurations. Deleting unsuccessful experiments can remove useful evidence about what was tested and why. Similarly, undocumented environment changes make comparisons difficult. Strong reproducibility practices improve collaboration, debugging, auditability, and confidence in model evaluation. They also make it easier to identify the exact configuration associated with a model that eventually moves toward production.

Question 398

What should an AI team do when a production incident reveals an unexpected model failure?

1. Immediately delete all evidence of the incident
2. Ignore the failure if average accuracy remains acceptable
3. Perform structured incident analysis and identify corrective actions
4. Retrain the model repeatedly without investigating the cause

Correct Answer: 3

Explanation

An unexpected production failure should trigger a structured incident-management process. The team should preserve relevant logs and evidence, assess the impact, identify the immediate cause and deeper contributing factors, and determine appropriate corrective actions. A root cause analysis can distinguish between issues involving data, model behavior, infrastructure, configuration, integration, human processes, or external changes. Corrective and preventive actions should then be documented and tracked. Simply retraining without understanding the failure can reproduce the same problem or introduce new risks. Incident findings can also inform monitoring improvements, testing requirements, governance controls, documentation, and future deployment decisions.

Question 399

Which concept describes the ability of an AI system to continue functioning appropriately when inputs contain reasonable levels of noise or variation?

1. Robustness
2. Data deletion
3. Label encoding
4. Model registration

Correct Answer: 1

Explanation

Robustness refers to an AI system’s ability to maintain acceptable behavior when exposed to expected variations, noise, or other changes in inputs. Robustness testing can involve altered data quality, environmental changes, small perturbations, unusual but plausible examples, or other conditions relevant to the intended operating environment. The objective is to identify situations in which model performance or behavior deteriorates unexpectedly. Robustness does not mean that a model will perform correctly under every possible condition. Instead, teams should define realistic operating boundaries and test representative challenges within those boundaries. Findings can guide model improvement, safeguards, monitoring thresholds, and human review requirements.

Question 400

What is an important activity when preparing an AI system for final retirement?

1. Increasing production traffic
2. Removing all historical documentation
3. Expanding the model’s permissions
4. Archiving required records and safely decommissioning the system

Correct Answer: 4

Explanation

AI retirement should be treated as a controlled lifecycle activity rather than simply turning off a production endpoint. The team should identify dependencies, communicate the retirement plan, preserve records required for governance or audit purposes, and safely remove or revoke unnecessary access. Relevant model, data, configuration, and decision records may need to be archived according to organizational retention requirements. Users and downstream systems should be transitioned to replacement processes where necessary. Security controls should also ensure that retired services cannot continue accessing sensitive resources. A structured retirement process reduces operational, compliance, and security risks while preserving useful knowledge for future AI initiatives.