Microsoft Certified Azure AI Fundamentals · Free Practice Question Medium
Question 28
A company has deployed an AI model for visual content analysis. During evaluation, they notice the model produces biased predictions due to an underrepresentation of certain groups in the training data. Which of the following strategies is best suited to mitigate this risk?
-
A
Switch to a prebuilt AI service, assuming it automatically eliminates bias.
-
B
Expand and diversify the training dataset and perform regular fairness audits.
-
C
Enhance the computational resources to speed up processing.
-
D
Increase the complexity of the model by adding more neural network layers.
Reveal correct answer
Correct answer: B
Explanation
Addressing Bias Through Data Diversity
When an AI model produces biased predictions due to the underrepresentation of certain groups in its training data, the root cause is an imbalanced dataset. Expanding and diversifying the training data is the most direct and effective approach to mitigating this risk. By including more representative samples of all groups, the model learns a more balanced view of the real world, which helps reduce bias in its predictions.
Regular Fairness Audits
Regular fairness audits are critical for ensuring that the model continues to perform equitably over time. These audits involve systematic checks of the model's outputs to detect any emerging biases and to monitor whether the training data remains representative. Such audits can also guide adjustments in data collection strategies and model retraining, ensuring that the model adapts to any shifts in the underlying population.
Why This Strategy Is Most Suitable
Root Cause Resolution:
Expanding and diversifying the dataset directly addresses the imbalance that is causing biased predictions. Without a representative dataset, any subsequent improvements in model architecture or computational power are unlikely to fully mitigate the bias.Proactive Monitoring:
Incorporating regular fairness audits creates a feedback loop, ensuring that bias is continually identified and addressed. This proactive approach prevents biases from becoming entrenched in the model over time.Long-Term Model Reliability:
By ensuring that the training data is balanced and that the model's performance is regularly evaluated, organizations can build AI systems that are not only more fair but also more reliable and trusted by all stakeholders.
Why Other Strategies Are Less Effective
Enhancing computational resources primarily improves processing speed and efficiency, but it does not correct for the underlying data imbalance that causes bias.
Switching to a prebuilt AI service might offer some level of bias mitigation, but these services are not guaranteed to automatically eliminate bias without proper oversight and additional context-specific adjustments.
Increasing the complexity of the model by adding more neural network layers generally leads to higher capacity but may also exacerbate overfitting and bias if the training data remains unbalanced.
Implementation Considerations
Data Collection:
Focus on gathering diverse and representative data that captures the full spectrum of relevant groups.Ongoing Monitoring:
Set up regular audits to assess model fairness and performance. Use these insights to inform periodic retraining and dataset updates.Ethical AI Guidelines:
Follow industry best practices and ethical guidelines to ensure that the AI model not only performs well technically but also meets societal standards for fairness and inclusivity.
Microsoft's Educational Content For This Topic:
A.
Although prebuilt AI services are designed for ease of use, they are not immune to bias issues. Responsible AI development still requires careful data and fairness considerations.
B.
Addressing bias effectively requires enhancing the diversity of the training data and conducting fairness audits to regularly monitor and mitigate skewed outcomes.
C.
While computational resources can affect speed and scale, they do not address data bias or fairness issues.
D.
Increasing model complexity does not resolve underlying data bias and may even worsen overfitting if the data isn’t representative.
Discussion
Think the marked answer is wrong, or have a better explanation? Share it below — comments appear after review.
