MLARM study plan

Define and contrast AI, machine learning, and generative AI, and classify ten example use cases by technique type

Map each core model risk (accuracy, overfitting, bias, drift, privacy leakage) to the lifecycle stage where it originates and the control that mitigates it

Explain the standard model risk management lifecycle stages (inventory, development, validation, approval, monitoring, decommissioning) in your own words

Draft a one-page risk taxonomy linking model failure modes to business, regulatory, and reputational consequences

Compare the scope and obligations of the major AI governance frameworks (NIST AI RMF, ISO/IEC 42001, OECD AI Principles, EU AI Act risk tiers) without conflating them

Classify five hypothetical AI systems into EU AI Act risk categories and justify each classification with the applicable obligation

Outline the roles of a three-lines-of-defense model in AI governance and assign responsibilities for each lifecycle stage

Identify the documentation artifacts a regulator would request during an AI supervisory review and note which framework mandates each

Distinguish data quality dimensions (completeness, accuracy, timeliness, consistency, representativeness) and design a check for each

Apply privacy principles (purpose limitation, minimization, retention, lawful basis) to a training-data pipeline and flag three design decisions that create privacy risk

Explain anonymization versus pseudonymization versus de-identification and identify where re-identification risk persists

Describe adversarial threats to ML systems (data poisoning, evasion, model extraction, membership inference) and select a defensive control for each

Map data-security controls to the data lifecycle from ingestion through model retirement

Write model documentation (intended use, limitations, assumptions, training data provenance) for a worked example model

Design a validation plan specifying conceptual soundness review, outcome analysis, benchmarking, and sensitivity testing

Distinguish developer, validator, and approver responsibilities and identify independence requirements

Construct a validation findings log with severity ratings and remediation owners

Select performance metrics appropriate to classification, regression, and ranking problems and explain why accuracy alone misleads on imbalanced data

Distinguish data drift, concept drift, and covariate shift, and choose a detection method for each

Design a monitoring plan with thresholds, alert routing, and escalation triggers for a production model

Explain challenger models, backtesting, and periodic revalidation cadence as drift-response mechanisms

Compare global and local explainability techniques (feature importance, SHAP, LIME, counterfactuals) and match each to an audience need

Define protected attributes, disparate impact, equal opportunity, and calibration, and compute a fairness metric on a small worked dataset

Diagnose sources of bias across data, labeling, feature selection, and deployment, and propose a mitigation for each

Explain the accuracy-fairness trade-off and how to document a justified trade-off decision

Assess third-party and vendor model risk, including due diligence, contract terms, audit rights, and concentration risk

Trace operational risk events across the model lifecycle and assign preventive and detective controls

Design model inventory fields and change-management triggers that force revalidation

Explain decommissioning criteria and the transition plan for retiring a model

Evaluate responsible AI principles (transparency, human oversight, accountability, sustainability) against a concrete deployment scenario

Summarize emerging topics (generative AI risks, agentic systems, model supply chain, AI incident reporting) and their governance implications

Synthesize all domains into an integrated review, reconciling overlapping obligations across frameworks

Complete two timed practice sets and analyze every missed item by domain and root cause