Risk Assessment & Management Flashcards

Which technique is most appropriate for quantifying uncertainty in a neural network's predictions to support risk-aware decision making?

Answer: Dropout at inference time (MC Dropout)

MC Dropout approximates Bayesian inference by running multiple stochastic forward passes at test time, yielding a distribution over predictions that reflects model uncertainty.

In a high-stakes ML deployment, what does 'model risk' primarily refer to?

Answer: Adverse outcomes arising from incorrect or misused model outputs

Model risk encompasses the potential for financial loss, regulatory penalty, or harm caused by errors in model development, implementation, or inappropriate use of predictions.

A credit scoring model shows strong performance on historical data but degrades rapidly after deployment. Which risk category best describes this phenomenon?

Answer: Data drift risk

Data drift occurs when the statistical properties of input features change over time post-deployment, causing models trained on historical distributions to underperform.

When assessing risks of an ML model in a regulated industry, which document type formally captures identified risks and corresponding mitigations?

Answer: A model risk register

A model risk register systematically catalogs identified risks, their likelihood, potential impact, and the mitigations or controls applied to each.

What is the primary purpose of stress-testing an ML model during risk assessment?

Answer: To evaluate model behavior under extreme or adversarial input conditions

Stress testing exposes the model to edge cases, outliers, and adversarial inputs to reveal failure modes that may not appear in standard evaluation.

Which risk mitigation strategy involves training an ML model on multiple diverse datasets to reduce vulnerability to any single distribution shift?

Answer: Ensemble diversification

Ensemble diversification trains models on varied data sources or with different architectures so that no single distribution shift degrades all ensemble members simultaneously.

In the context of ML risk management, 'operational risk' most directly refers to:

Answer: Failures in people, processes, or systems supporting the model's operation

Operational risk in ML covers failures stemming from inadequate internal processes, human error, system failures, or external events affecting the model's production pipeline.