Google Professional Machine Learning Engineer Question 474
Select 3Google Cloud PlatformYou are building an AI solution for a financial institution to automatically approve or reject loan applications. During deployment readiness checks, you want to ensure the model is fair and does not exhibit bias against any demographic groups. Which of the following steps should you take to assess your AI solution's readiness?
- A
Perform a fairness audit by testing model output across different demographic categories, such as age, gender, and race.
- B
Use Explainable AI tools to evaluate feature importance and detect whether sensitive attributes are influencing predictions disproportionately.
- C
Ensure the training dataset includes an equal number of samples from all demographic groups to avoid bias.
- D
Conduct a human-in-the-loop review of the model's decisions, focusing on edge cases and high-risk predictions.
- E
Run adversarial testing to determine whether the model is susceptible to manipulation or unfair outcomes.
Show answer and explanation
Correct answers: A, B, D
Explanation
Assessing AI solution readiness for fairness and bias involves auditing the model's outputs, using Explainable AI tools to understand prediction drivers, and incorporating human oversight for high-risk scenarios. These steps ensure that the model meets ethical and fairness standards, while adversarial testing and dataset balancing, although useful, are not directly sufficient for this purpose.
- A. Correct.
Performing a fairness audit is a critical step to assess whether the model produces biased outcomes for specific demographic groups and aligns with fairness requirements.
- B. Correct.
Using Explainable AI tools helps identify which features are driving predictions. This ensures that sensitive attributes, like race or gender, are not unfairly influencing decisions.
- C. Incorrect.
While having balanced data is important, equal representation alone does not guarantee fairness, as bias can still arise from systemic patterns in the data or the model's design.
- D. Correct.
Human-in-the-loop review is a vital step for assessing edge cases and ensuring that high-stakes decisions meet ethical and fairness standards.
- E. Incorrect.
Adversarial testing focuses on robustness and security, not fairness. While useful for other assessments, it is not directly relevant to fairness or bias evaluation.