PECB Certified Artificial Intelligence Professional cheat sheet

Supervised learning maps labelled input-output pairs to a function; unsupervised learning finds structure in unlabelled data; reinforcement learning optimises a policy via reward signals from an environment.

Bias-variance decomposition: expected error = bias² + variance + irreducible noise; high bias underfits (simple model), high variance overfits (complex model).

Precision = TP/(TP+FP); Recall = TP/(TP+FN); F1 = 2·(Precision·Recall)/(Precision+Recall); use F1 or PR-AUC when classes are imbalanced, accuracy is misleading.

Confusion-matrix accuracy = (TP+TN)/(TP+TN+FP+FN); a model predicting the majority class on 99:1 data scores 99% accuracy while recall on the minority class is 0.

Regularisation decision rule: L2 (ridge) shrinks coefficients and stabilises correlated features; L1 (lasso) drives coefficients to zero and performs feature selection.

Cross-validation: k-fold trains k models on k−1 folds and validates on the held-out fold; the test set is touched once, after model selection, never during tuning.

Neural network training: forward pass computes loss, backpropagation computes gradients via the chain rule, an optimiser (SGD, Adam) updates weights by gradient × learning rate.

Transformer attention: Attention(Q,K,V) = softmax(QKᵀ/√d_k)V; self-attention lets each token weight all others, removing the sequential dependency of RNNs.

Generative model families: GANs pit generator against discriminator; VAEs optimise an evidence lower bound; diffusion models learn to reverse a noise-adding process.

ISO/IEC 23894 structures AI risk management on the ISO 31000 cycle: establish context, identify, analyse, evaluate, treat, monitor, communicate - applied to AI-specific harms.

NIST AI RMF has four functions: GOVERN (culture, accountability), MAP (context and risks), MEASURE (metrics and testing), MANAGE (prioritise, respond, recover).

Risk = likelihood × impact; treatment options are mitigate, transfer, avoid, or accept, and residual risk after treatment must be recorded with an owner and acceptance date.

NIST AI RMF trustworthiness characteristics: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, fair with harmful bias managed.

Bias sources by stage: historical bias in training data, representation bias in sampling, measurement bias in labelling, and deployment bias when a model is used outside its validation population.

Model validation evidence expected: holdout or out-of-time performance, subgroup metrics (disaggregated by protected attribute), stability over time, and documented limitations.

Human oversight design rule: keep a human in the loop for high-impact decisions, with authority, competence, and time to override the system, not a nominal rubber-stamp.

Drift taxonomy: data drift (input distribution shifts), concept drift (input-output relationship shifts), and label drift; each requires monitoring thresholds and retraining triggers.

OECD AI Principles (2019, updated 2024): inclusive growth and sustainable development, human-centred values, transparency and explainability, robustness and safety, accountability.

EU AI Act risk tiers: prohibited practices, high-risk systems (Annex III areas such as employment, credit, education, critical infrastructure), limited-risk with transparency duties, minimal-risk.

EU AI Act general-purpose AI obligations: transparency and copyright policy for all GPAI providers; systemic-risk models additionally require model evaluation, incident reporting, and cybersecurity.

ISO/IEC 42001 is a certifiable AI management system standard following Plan-Do-Check-Act, with clauses on leadership, risk assessment, operational control, and performance evaluation.

AI governance roles: governing body sets risk appetite, an AI risk committee or owner approves use cases, model owners are accountable, and independent audit or second-line review verifies.

AI inventory minimum fields: use case, business owner, model type and version, data sources and sensitivity, affected population, risk rating, human-oversight level, review date.

Transparency duties: disclose that a user is interacting with an AI system, label synthetic media where required, and provide meaningful information about logic and consequences of outputs.

Digital transformation governance: benefits realisation is measured against baseline KPIs, and portfolio stage gates (ideation, pilot, scale, retire) control investment decisions.

Adversarial attack classes: evasion (crafted input at inference), poisoning (corrupted training data), model extraction (query-based stealing), and membership inference (inferring training-set membership).

Prompt injection: direct injection comes from user input, indirect injection is embedded in retrieved documents or tool output; mitigations include input/output filtering and least-privilege tool access.

Data protection principles (GDPR Art. 5): lawfulness, fairness and transparency; purpose limitation; data minimisation; accuracy; storage limitation; integrity and confidentiality; accountability.

GDPR Art. 22 restricts solely automated decisions with legal or similarly significant effects; safeguards include human intervention, expressing a view, and contesting the decision.

DPIA (GDPR Art. 35) is required for high-risk processing such as large-scale profiling or systematic monitoring; AI systems using sensitive data typically trigger it.

Privacy-enhancing technologies: differential privacy adds calibrated noise with a privacy budget ε (lower ε = stronger privacy), federated learning keeps data local, secure multi-party computation and homomorphic encryption protect data in use.

Defence in depth for ML pipelines: data provenance and integrity checks, access control and secrets management, model artefact signing, adversarial testing, and runtime anomaly monitoring.

OWASP LLM Top 10 categories include prompt injection, insecure output handling, training-data poisoning, model denial of service, supply-chain risks, sensitive information disclosure, and excessive agency.

Incident response for AI: detect via monitoring, contain (disable model or revoke keys), eradicate the cause, recover, and notify regulators within statutory deadlines (e.g. 72 hours under GDPR).