ML in Practice: Bias, Fairness, and Ethics Quiz Quiz

Explore key concepts related to machine learning bias, fairness, and ethical considerations with this quiz. Deepen your understanding of responsible ML deployment, common pitfalls, and best practices to ensure equitable and ethical outcomes.

1. Recognizing Bias in Machine Learning

   Which scenario best demonstrates selection bias in a machine learning dataset intended to predict housing prices?

   1. Data is normalized incorrectly, causing high variance.
   2. The model uses too many hidden layers, causing overfitting.
   3. The dataset has unbalanced numerical ranges for features.
   4. The dataset only includes urban property listings and excludes rural areas.

   Explanation: Selection bias occurs when the data sampled is not representative of the desired population, such as only including urban property listings while excluding rural areas. This skews the model's predictions and reduces fairness. Overfitting due to too many hidden layers is a training issue, not a data bias. Incorrect normalization affects feature scaling, not selection bias. Unbalanced numerical ranges might impact model performance but do not constitute selection bias.

2. Fairness-Aware ML Strategies

   Which approach is most effective in reducing algorithmic discrimination when training a machine learning model for loan approvals?

   1. Labeling outliers as special classes for separate processing.
   2. Removing sensitive attributes like race or gender from training data.
   3. Gathering larger datasets without considering attribute distribution.
   4. Increasing model complexity with deeper neural networks.

   Explanation: Eliminating sensitive attributes like race or gender can help reduce direct discrimination and make the model fairer. Simply increasing model complexity does not address fairness concerns. Gathering larger datasets is beneficial in general but without attention to attribute distribution might perpetuate existing biases. Treating outliers separately addresses rare cases, not discrimination against protected groups.

3. Ethical Principles in ML Deployment

   Why is transparency important when deploying machine learning applications in healthcare decision-making?

   1. Transparency allows for quicker hardware acceleration.
   2. Transparency ensures all input features are binary values.
   3. Transparency improves model precision by increasing neuron activations.
   4. Transparency helps stakeholders understand and trust model recommendations.

   Explanation: Transparency leads to better understanding and trust of ML model decisions by clinicians and patients, which is critical in healthcare. Improving model precision is unrelated to transparency. Hardware acceleration is a technical efficiency concern, not an ethical one. Input feature types are separate from the need for transparent explanations.

4. Consequences of Data Bias

   What is a likely outcome if a computer vision system for facial recognition is trained primarily on images of adults but rarely on those of children?

   1. The system will overfit on all test samples equally.
   2. The system benefits from increased robustness on unknown samples.
   3. The system may show reduced accuracy in identifying children’s faces.
   4. The system will automatically normalize to all age groups.

   Explanation: If training data underrepresents children, the model may not generalize well and thus misidentify or inaccurately recognize children’s faces. Overfitting does not necessarily mean balanced error across all groups. Increased robustness would require diverse data, not a lack of it. Automatic normalization does not compensate for insufficient representation of specific groups.

5. Evaluating Fairness Metrics

   Which fairness metric compares the rate of positive outcomes between different demographic groups in a binary classification problem?

   1. Mean absolute error
   2. Root mean square error
   3. Area under the ROC curve
   4. Demographic parity

   Explanation: Demographic parity measures fairness by comparing the rates of favorable outcomes among groups, ensuring no group is unduly advantaged. Mean absolute error and root mean square error are prediction accuracy metrics, not fairness-related. The area under the ROC curve assesses overall model discrimination ability, not parity between groups.