Select Two (Multi-Select) Questions

Hyperparameter tuning

Why wrong: Hyperparameter tuning is part of model optimization.

Model evaluation

Why wrong: Model evaluation is after training.

One-hot encoding all features

Why wrong: Incorrect; one-hot encoding is only for categorical features, and not all features require it.

Normalizing features

Why wrong: Incorrect; normalization is beneficial but not essential for all algorithms.

Removing outliers

Why wrong: Incorrect; outlier removal is optional and depends on the problem.

F1 score

Why wrong: F1 is useful but the question asks for two; precision and recall are more direct.

Area under the ROC curve (AUC-ROC)

Why wrong: AUC-ROC is less informative for imbalanced datasets; precision-recall is preferred.

Accuracy

Why wrong: Accuracy is high even if the model predicts all non-fraud, so it's misleading.

Using adversarial training

Why wrong: Adversarial training improves robustness to adversarial examples, not bias.

Applying L1 regularization

Why wrong: L1 regularization induces sparsity, not fairness.

Performing k-fold cross-validation

Why wrong: Cross-validation assesses performance, does not mitigate bias.

Gradient boosting

Why wrong: Gradient boosting is a specific algorithm, not a learning paradigm.

Quantum computing

Why wrong: Quantum computing is a computing paradigm, not a machine learning paradigm.

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: Increased complexity can make the model more sensitive to small perturbations, making attacks easier.

Hyperparameter optimization using grid search.

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

Principal Component Analysis (PCA)

Why wrong: PCA is feature extraction, not selection.

SMOTE

Why wrong: SMOTE is for class imbalance.

Dropout

Why wrong: Dropout is for regularization, not feature selection.

Batch size.

Why wrong: Batch size affects gradient noise, not capacity directly.

Optimizer choice (e.g., SGD vs Adam).

Why wrong: Optimizer affects training dynamics, not capacity.

Learning rate.

Why wrong: Learning rate does not change model capacity.

Use larger convolutional filters (e.g., 7x7 instead of 3x3) to capture more context.

Why wrong: Larger filters increase computation and slow down inference.

Increase the number of convolutional layers to improve feature extraction.

Why wrong: More layers increase computational cost and latency.

Encrypting the model weights

Why wrong: Encryption protects the model at rest, not during inference.

Continuous model monitoring

Why wrong: Monitoring detects attacks but does not prevent them.

Rate limiting API access

Why wrong: Rate limiting reduces abuse but does not defend against crafted inputs.

Perform stemming or lemmatization.

Why wrong: Stemming is done after tokenization.

Remove stop words from the text.

Why wrong: Stop word removal is done after tokenization.

Replace missing values with a placeholder.

Why wrong: Missing value handling is not specific to preprocessing before tokenization.

Boosting

Why wrong: Boosting combines weak learners and may overfit if not carefully tuned.

Increasing the learning rate

Why wrong: Learning rate affects convergence, not overfitting directly.

Increasing the number of hidden layers

Why wrong: More layers increase model complexity, potentially worsening overfitting.

Basing acceptance solely on training accuracy

Why wrong: Training accuracy does not reflect real-world performance.

Maximizing model complexity to achieve the best accuracy

Why wrong: Complexity can lead to maintenance issues and overfitting.

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.

Logistic regression

Why wrong: Incorrect; logistic regression is for binary classification.

K-means

Why wrong: Incorrect; k-means is for clustering, not regression.

Use Git to version model binaries.

Why wrong: Git is designed for code, not large binary files.

Keep only the latest model to save storage.

Why wrong: Versioning requires keeping historical models for comparison and rollback.

Hyperparameter tuning

Why wrong: Hyperparameter tuning is performed during model development, not continuously in production.

Feature importance analysis

Why wrong: Feature importance is used for interpretability, not directly for maintaining performance.

Schedule automatic weekly retraining of the model.

Why wrong: Retraining is a remediation action, not a monitoring method.

Increase the model's complexity to improve generalization.

Why wrong: Increasing complexity is not a monitoring practice and may cause overfitting.

Monitor the average prediction confidence of the model.

Why wrong: Prediction confidence may not indicate drift; it can be high even on drifted data.

Maximizing cost reduction for the hospital

Why wrong: Cost reduction is not an ethical priority.

Replacing human judgment entirely with AI

Why wrong: AI is meant to assist, not replace.

Using open-source models to reduce licensing costs

Why wrong: Licensing is not an ethical concern.

Ignore missing values and train the model directly.

Why wrong: Most models cannot handle missing values directly.

Impute missing values with the mean of the entire dataset.

Why wrong: Mean imputation can distort distributions and is not always appropriate.

Replace missing values with the most frequent value always.

Why wrong: Replacing with mode is not universally appropriate and may introduce bias.