Select Two (Multi-Select) Questions

Increase the complexity of the model to capture more patterns

Why wrong: Increasing model complexity does not address the measured selection-rate disparity; it risks amplifying existing biases by fitting noise rather than enforcing fairness constraints. This option is tempting because adding complexity can improve predictive accuracy on imbalanced data, which would be correct if the goal were reducing underfitting rather than mitigating demographic bias.

Add more training data from both groups

Why wrong: Adding more data without balancing may not reduce bias.

Remove the protected attribute from the model input

Why wrong: Removing the attribute may not eliminate bias due to proxy features.

Using time-based splitting for sequential data

Why wrong: Time-based splitting is designed to prevent leakage by respecting temporal order.

Using cross-validation on the entire dataset

Why wrong: Cross-validation is a proper validation technique and does not cause leakage when done correctly.

Model pruning to reduce the number of parameters.

Why wrong: Pruning can remove neurons that are critical for detecting adversarial perturbations, reducing robustness.

Increasing model complexity to capture more patterns.

Why wrong: Increasing model complexity enlarges the attack surface by introducing more parameters for gradient-based adversarial perturbations to exploit, directly contradicting the security requirement for robustness. This option is tempting because deeper models often improve accuracy on benign data, making it a correct choice for performance optimisation tasks where generalisation, not adversarial defence, is the primary goal.

Hyperparameter optimization using grid search.

Why wrong: Hyperparameter optimization does not directly defend against adversarial attacks.

Feature importance scores

Why wrong: High importance does not directly indicate bias.

Confusion matrix on the entire dataset

Why wrong: Without subgroup breakdown, bias may be hidden.

Cross-validation accuracy

Why wrong: Accuracy alone does not indicate bias.

Ignore the disparity because the model is accurate overall

Why wrong: Ignoring disparity may violate anti-discrimination laws.

Remove all demographic attributes from the dataset

Why wrong: Removing attributes may not eliminate bias due to proxy variables, and can hinder fairness analysis.

Shadow deployment

Why wrong: Shadow deployment duplicates traffic to the new version but does not serve user requests, so it does not test real behavior.

Direct cutover

Why wrong: Direct cutover replaces the old version instantly with no testing period.

Blue-Green deployment

Why wrong: Blue-Green switches all traffic at once, not gradual testing.

A/B testing framework for comparing models

Why wrong: A/B testing is part of deployment, not CI/CD pipeline design.

Automated unit testing of application code

Why wrong: Unit testing is important but not specific to ML pipelines.

The model needs to understand complex legal terminology

Why wrong: Fine-tuning is better for embedding domain-specific language; RAG relies on retrieval quality.

The queries require deep reasoning across multiple documents

Why wrong: Both approaches can handle multi-document reasoning; fine-tuning may even be better for integrating cross-document knowledge.

The company has limited compute budget for training

Why wrong: RAG may require more inference compute (retrieval + generation), but fine-tuning has high upfront training cost.

Using a larger learning rate

Why wrong: Larger learning rates can cause divergence, not prevent overfitting.

L1 regularization

Why wrong: L1 regularization is used but not as common as dropout and early stopping for overfitting.

Increasing the number of layers

Why wrong: Adding layers often increases overfitting.

Deploy monitoring for anomalous input patterns

Why wrong: Monitoring detects attacks but does not prevent the model from performing excessive actions; it is a detective control, not a preventative one.

Apply output filtering to block sensitive data in responses

Why wrong: Output filtering prevents data leakage but does not directly mitigate the risk of the model taking unauthorized actions (excessive agency).

Supervised learning is only used for classification

Why wrong: Supervised learning includes regression as well.

Unsupervised learning always requires a target variable

Why wrong: Unsupervised learning has no target variable.

Supervised learning is a subset of reinforcement learning

Why wrong: They are distinct paradigms.

Reduce the number of features using principal component analysis (PCA).

Why wrong: PCA reduces dimensionality but does not address class imbalance.

Use accuracy as the primary evaluation metric.

Why wrong: Accuracy can be misleading in imbalanced datasets because a model that predicts all majority class can achieve high accuracy.

Explicit consent from all data subjects

Why wrong: Consent is not always required; other legal bases may apply.

Data retention period of at least 10 years

Why wrong: Retention periods vary; no universal 10-year requirement.

Cloud-based deployment for scalability.

Why wrong: Deployment method is operational, not governance.

Model accuracy thresholds for production deployment.

Why wrong: Accuracy is a performance metric, not a governance component.

Increasing training data

Why wrong: While more data can help, it is not a 'technique' in the same sense; it is a data collection strategy, not a regularization technique applied during training.

Reducing number of layers

Why wrong: Reducing layers can reduce overfitting but also reduces model capacity significantly and may lead to underfitting; it is not a primary regularization technique.

Cross-validation

Why wrong: Cross-validation is a method to evaluate models, not a cause of bias.

High regularization

Why wrong: Regularization reduces variance, not bias.

Increase the number of training epochs to ensure convergence.

Why wrong: Epoch count affects the model, but does not guarantee reproducibility.

Use the same GPU hardware for both training and inference.

Why wrong: GPU hardware differences may cause numerical variations.

Use parallel data loading to speed up inference.

Why wrong: Parallelism can introduce non-deterministic ordering.

Adding Gaussian noise

Why wrong: Noise can help but is less common than geometric and color augmentations.

Random cropping

Why wrong: Cropping can help but may cut out important features; not always best.

Transformer encoder only

Why wrong: A Transformer encoder alone lacks a temporal modelling mechanism, such as recurrent connections or attention over time, which is required to track movement patterns across frames in a video stream. It processes each frame independently, so it cannot encode the sequential dependencies needed for motion analysis. This option is tempting because Transformer encoders excel at spatial feature extraction in static images, making them a correct choice for single-frame object detection tasks.

Generative adversarial network (GAN)

Why wrong: GANs are for generation, not detection or tracking.

Variational autoencoder (VAE)

Why wrong: VAEs are also for generation, not detection.

Training speed

Why wrong: Incorrect; training speed is an efficiency metric, not ethical.

Model accuracy

Why wrong: Incorrect; accuracy is a performance metric, not an ethical consideration.

Model size

Why wrong: Incorrect; model size is a resource consideration, not ethical.