Domain 3 - AI Algorithms and Models

MCQ
1. How can AI developers ensure they comply with regulatory requirements? Choose 3
answers.
Store data in multiple locations
Have metrics and tools in place that will analyze system performance concerning regulatory
compliance
Put one person in charge of managing regulatory compliance
Understand what the regulatory requirements are
Define thresholds and requirements for AI systems relevant to regulatory requirements
2. What are three factors to consider when evaluating an AI model?
Trends
Speed
Cost
Outcomes
Size
3. to best evaluate a model, a separate data set from what is used for training should be used.
True
False
4. What is an algorithm?
The relative change of a model and its output
An ordered list of numerical properties of observed phenomena
A procedure with a sequence of operations
Data gathered on an AI model and used to ensure that an AI model is working properly
5. Sensitivity refers to how well a model identifies negative instances, while specificity refers to
how well a model identifies positive instances.
True
False
6. What is a self-fulfilling prophecy?
An algorithm that identifies and removes existing patterns in data
An algorithm that reinforces and magnifies existing patterns in data
An algorithm that only looks for and uses data that supports an existing idea
An algorithm that uses data collected from samples that do not accurately represent the
population
7. New testing data should be introduced to a model as-is to test robustness.
True
False
8. Stakeholder approval often stems from AI developers testing results that are easy to
understand.
True
False
9. What is an iteration?
 A series of steps used to transform data used for an AI model into a final prediction or
outcome
A chart or report that explains the results of an AI model or the data used to build it
A sequence of retraining a model using different settings, such as different parameters
Data that is paired with output targets in an AI model
10. What is specifity?
The ability of an algorithm to classify data accurately
The relative change of a model’s output when data is input
The ability of an algorithm to identify nonrelevant outputs accurately
The relative change of a model’s output when the data set is changed
11. What is the purpose of using new data to test a model?
To create baseline metrics for the model
To determine trends in the data outputted by the model
To verify the performance of the AI model
To confirm that all data is formatted correctly
12. What are two ways to use trends while evaluating data?
Confirming adherence to regulations
Evaluating algorithmic families
Making predictions
Detecting anomalies
Testing the specificity of a model
13. What does accuracy measure?
The proportion of correctly predicted negative instances out of all instances with negative
predictions
The proportion of correctly predicted positive instances out of all positive instances
The proportion of correctly predicted positive instances out of all instances with positive
predictions
The overall correctness of a model’s predictions
14. Why should performance metrics be gathered on an AI model?
To determine which algorithm is being used
To demonstrate the limitations of the model
To promote transparency and ethical use of an AI system
To ensure it is working properly
15. Match the algorithm family to its description.
Reinforcement learning  learns from interacting with its environment and receiving
rewards or penalties.
Unsupervised learning  works with unlabeled data to uncover unknown, hidden patterns
or structures in the data.
Deep learning  uses neural networks with AI models for human based interactions.
Supervised learning  use this labeled data where a target variable is known.
16. What is sensitivity?
The relative change of a model’s output when data is input
The ability of an algorithm to accurately classify data
The ability of an algorithm to accurately identify nonrelevant outputs
The relative change of a model’s output when the data set is changed
17. What is the purpose of data visualizations?
To clean and sanitize data that will be used to train a model
To summarize data in an easily understandable manner
To establish key performance indications
To estimate the cost of an AI model
18. What is an explainable algorithm?
Algorithm inputs that are connected to AI model training data
Algorithm inputs that are connected to real world populations
Algorithm results that are connected to algorithm inputs rather than extraneous or
incorrectly interpreted features
Algorithm results that are connected to extraneous or incorrectly interpreted features rather
than algorithm inputs
19. What should be included when documenting regulatory requirements? Choose 4 answers
Amount of stored data
Evaluation methods
Compliance measures
Security procedures
Corrective actions for noncompliance
Financial cost
20. What factor is being evaluated when an organization needs a cloud to build its AI model?
Speed
Cost
Trends
Outcomes it is producing
21. Which statement best defines a stakeholder?
And individual, group, or organization that may affect, the affected by, or perceive itself to
be affected by a project.
An individual that perceives themselves to be affected by a project.
An individual, group, or organization that may affect a project.
An individual, group, or organization that may be affected by a project.
22. How can one reduce imbalances in data?
Remove certain data
Replace certain data
Collect more data
Create multiple instances of certain data within a dataset
23. When does underfitting occur?
When a model does not capture underlying patterns in data and thus performs poorly on
training data or new, unseen data.
When a model generates a large number of false positives.
When a model learns training data very well but performs poorly on new, unseen data.
When a small change in the input data leads to a significant change in the output.
24. Match the algorithm to its definition.
Decision tree algorithm  a supervised learning method in which samples are split into two
or more sets based on input variable differentiators.
 Neural network  an algorithm designed to mimic the human brain that uses artificial
neurons to recognize and obtain information.
K means clustering  and unsupervised learning method that helps split data into an
unspecified number of groups.
25. What should happen during session meetings for evaluating AI solutions? Choose 3 answers.
Define objectives for the session
have materials ready that describe the AI solution
ensure there is an opportunity for discussion
ensure every attendee understands all parts of the algorithms used in the project
begin implementing at least one AI solution
26. Which 2 facts about visualizations are true?
Any visualization created from AI data will be useful in making business decisions.
Visualizations are used to verify the specificity of a model.
The precision of a model can be determined using visualizations.
Visualizations should help determine whether an AI model is accurate.
One needs to understand the criteria of the decisions to be made when assessing visuals.
27. How should AI developers handle instances of noncompliance? Choose 3 answers.
Perform fewer audits of logged activity.
Take additional security measures to ensure an AI project is within a regulatory bodies’
frameworks.
Wait for regulatory bodies to contact them regarding noncompliance mitigation.
Refine the AI model.
Refine how data is processed and collected.
28. Which three statements regarding explainability requirements are true?
They induce interpretability so people can understand AI model results.
They are a series of steps used to transform data used for an AI model into a final prediction
or outcome.
They promote transparency and trust in AI systems.
They are clear explanations for decisions, protections, and recommendations made from an
AI model.
They include role based access control (RBAC)
29. What happens if an AI model does not perform as anticipated?
The model must be used as-is.
Parameters can be adjusted and a model can be retrained.
A new AI model must be built.
A different algorithm must be chosen.
30. What should happen first in session meetings for evaluating AI solutions?
Outline the budget.
Identify sensitive data.
Outline non-AI solutions.
Identify stakeholders.
31. How can AI developers reduce the risk of self-fulfilling prophecies affecting AI outcomes?
Choose 2 answers.
Consider where the data is stored.
Consider historical biases.
 Consider the age of the AI model.
Consider how the training data compares to real-world data.
Consider how the AI model can impact the data used in the algorithm.
32. A model can be trained using multiple training iterations but the same parameter
combinations must be used.
True
False
33. How can one prevent bias in real world data input into an AI model?
Manually enter data rather than uploading data.
Only use data collected from certain demographics.
Compare real world inputs to training data.
Implement an algorithm that removes bias.