NCA-GENM Question 225
Select 3A company is developing a generative AI model for multimodal tasks, including image and text generation. During testing, they notice that the model's outputs exhibit bias against certain demographic groups. Which of the following steps should the company prioritize to minimize bias in their AI system?
- A
Ensure the training dataset is representative of diverse demographic groups.
- B
Use data augmentation techniques to artificially increase the presence of underrepresented groups in the dataset.
- C
Avoid pre-processing the dataset to preserve its original form and maintain authenticity.
- D
Implement fairness-aware algorithms during model training to reduce bias.
- E
Rely solely on post-processing techniques to modify biased outputs after generation.
Show answer and explanation
Correct answers: A, B, D
Explanation
Minimizing bias in AI systems requires addressing bias at multiple stages, including data preparation, model training, and evaluation. Ensuring a representative dataset, using data augmentation to address imbalances, and applying fairness-aware algorithms are proactive, effective strategies. Solely relying on post-processing or avoiding dataset pre-processing does not address the root causes of bias and may lead to suboptimal outcomes.
- A. Correct.
Ensuring the training dataset is representative of diverse demographic groups is a fundamental step in minimizing bias, as biased data directly impacts the model's output.
- B. Correct.
Using data augmentation to increase the presence of underrepresented groups can help balance the dataset and reduce bias during training.
- C. Incorrect.
Avoiding pre-processing the dataset is not a good practice in this context, as pre-processing can help remove existing biases or anomalies in the data.
- D. Correct.
Fairness-aware algorithms are designed to mitigate bias during model training and are a proactive approach to addressing bias at the model level.
- E. Incorrect.
Relying solely on post-processing techniques is insufficient, as it only addresses the symptoms of bias rather than the root causes in the data or training process.