HIRAF cheat sheet

AI is the field of building systems that perform tasks associated with human intelligence: perception, reasoning, learning, language, and action.

Symbolic (rule-based) AI reasons with explicit knowledge and logic; connectionist (subsymbolic) AI learns statistical patterns in weights, as in neural networks.

Supervised learning uses labeled input-output pairs; unsupervised learning finds structure in unlabeled data; reinforcement learning optimizes actions via reward signals from an environment.

A machine learning model is a parameterized function mapping inputs to outputs; training adjusts parameters to minimize a loss function over training data.

Generalization is performance on unseen data; overfitting is low training error with high test error, underfitting is high error on both.

Strong (general) AI denotes human-level breadth across tasks; narrow AI is task-specific and describes all deployed production systems today.

The Turing test operationalizes intelligence as indistinguishable conversational behavior; it measures behavior, not internal understanding.

Intelligence tests of capability (benchmarks) measure task accuracy; they do not by themselves establish robustness, fairness, or safety.

Bias-variance tradeoff: high bias yields underfitting; high variance yields overfitting; total error = bias² + variance + irreducible noise.

Regularization (L1/Lasso, L2/Ridge) penalizes parameter magnitude; L1 drives sparsity, L2 shrinks weights smoothly.

Cross-validation (k-fold) estimates generalization by rotating held-out folds; the test set is touched only once, at the end.

Classification metrics: precision = TP/(TP+FP); recall = TP/(TP+FN); F1 = harmonic mean of precision and recall.

ROC-AUC measures ranking quality across thresholds; PR-AUC is preferred under heavy class imbalance.

Data leakage occurs when information from the target or test set contaminates training features, inflating validation scores.

Hyperparameters (learning rate, depth, regularization strength) are set before training; parameters are learned during training.

Feature engineering creates informative inputs; feature scaling (standardization, normalization) matters for distance- and gradient-based models.

Model selection compares candidates on validation data; final reported performance comes from an untouched holdout or nested cross-validation.

Data governance defines decision rights and accountability for data quality, access, privacy, and lifecycle across an organization.

Data quality dimensions: accuracy, completeness, consistency, timeliness, validity, and uniqueness.

Metadata is data about data (schema, lineage, ownership); data lineage traces origin and transformation of each field.

Data minimization: collect and retain only data necessary for a stated purpose; retention limits reduce exposure and compliance risk.

Personal data is any information relating to an identified or identifiable natural person; anonymization must be irreversible to fall outside scope.

Master data management maintains a single trusted reference copy of core entities (customer, product) across systems.

Role separation: data owner (accountable), data steward (quality operations), data custodian (technical controls).

Data classification tiers (public, internal, confidential, restricted) drive access controls, encryption, and retention rules.

Core AI risk categories: performance/robustness, privacy, security, bias and fairness, transparency, accountability, and misuse.

Fairness definitions conflict: demographic parity, equalized odds, and individual fairness cannot generally all hold simultaneously.

Proxy variables (postal code, name) can reproduce protected-attribute discrimination even when that attribute is excluded.

Robustness covers distribution shift, adversarial inputs, and data drift; monitoring and stress testing are the standard controls.

Explainability spans intrinsic (interpretable models) and post hoc (SHAP, LIME, counterfactuals) methods; explanation fidelity must be validated.

Human oversight: human-in-the-loop (approval), human-on-the-loop (monitoring), and full automation are selected by risk and reversibility.

Risk treatment options are avoid, mitigate, transfer, and accept; residual risk must be documented and owned.

Ethical principles commonly cited: beneficence, non-maleficence, autonomy, justice, explicability, and privacy protection.

Governance structures: AI policy, roles (AI officer, ethics board), risk-tiering, model inventory, and audit trails.

The EU AI Act uses risk tiers: prohibited practices, high-risk systems, limited-risk (transparency duties), and minimal-risk.

High-risk AI obligations include risk management, data governance, technical documentation, logging, human oversight, accuracy, robustness, and cybersecurity.

GDPR principles: lawfulness, fairness, transparency, purpose limitation, data minimization, accuracy, storage limitation, integrity, accountability.

GDPR data-subject rights include access, rectification, erasure, restriction, portability, and objection, including to automated decision-making.

Automated decision-making with legal or significant effects requires safeguards such as human intervention and the right to contest.

NIST AI RMF functions: Govern, Map, Measure, Manage; ISO/IEC 42001 specifies requirements for an AI management system.

Model documentation artifacts include model cards (intended use, metrics, limits) and system cards/datasheets for datasets.

MLOps lifecycle: data ingestion, feature pipeline, training, evaluation, deployment, monitoring, and retraining.

Deployment patterns: batch, real-time API, streaming, and edge; choice depends on latency, volume, and connectivity constraints.

Model drift types: data drift (input distribution change), concept drift (input-output relationship change), and label drift.

Monitoring signals: prediction distribution, input feature distributions, latency, error rates, data quality, and business KPIs.

Shadow deployment runs a new model alongside production without serving its output; canary release serves it to a small traffic slice.

Versioning must cover code, data, model artifacts, and configuration to make results reproducible and rollback-safe.

Feature stores centralize feature computation for training and serving to prevent training-serving skew.

Incident response for AI includes detection, rollback or kill switch, root-cause analysis, and post-incident documentation.

Serving infrastructure must enforce input validation, rate limits, access control, and secrets management.