AWS Certified AI PractitionerResponsible AIMedium
A healthcare provider is implementing an AI system to assist with disease diagnosis. During development, it's discovered that the model performs significantly worse on data from a particular ethnic minority group due to underrepresentation in the training dataset. Which of the following is the most appropriate best practice to address this issue?
- AImplement differential privacy techniques during model training.
- BCollect additional, representative data for the underrepresented group and retrain the model.
- CIncrease the model's overall accuracy by adding more data from the majority population.
- DRemove the ethnic minority group's data entirely from the training set to avoid bias.
Show answer & explanationAnswer & explanation
Correct answer: B. Collect additional, representative data for the underrepresented group and retrain the model.
The most effective way to address poor performance due to underrepresentation is to gather more representative data for that specific group and retrain the model, directly improving its ability to generalize to that population.
Why the other options are wrong
- A. Differential privacy helps protect individual data points but doesn't solve the issue of insufficient or unrepresentative data for a specific group.
- C. Increasing majority data might improve overall accuracy but would likely exacerbate the bias against the underrepresented group.
- D. Removing the data would prevent the model from learning anything about that group, making it unable to diagnose patients from that group at all, which is a worse outcome.
Addressing Data Bias
Addressing data bias involves identifying and mitigating systematic errors or imbalances in training data that can lead to unfair or inaccurate AI model predictions.
- Data quality is paramount for fair AI.
- Underrepresentation leads to poor performance on specific groups.
- Solutions include data augmentation, re-sampling, and collecting new data.
Memory trick: DATA CLEANUP for AI Trust