PECB Certified Artificial Intelligence Professional flashcards

33 cards. Read the term and the definition here. The deck at the top of this page is optional.

Supervised learning

A training paradigm in which a model learns a mapping from inputs to labelled outputs, minimising a loss between predictions and known targets.

Bias-variance tradeoff

Expected prediction error decomposes into bias squared, variance, and irreducible noise. Simpler models raise bias; more complex models raise variance.

Precision and recall

Precision is true positives divided by predicted positives. Recall is true positives divided by actual positives. They trade off as the decision threshold moves.

L1 versus L2 regularisation

L1 adds the absolute value of weights to the loss and drives some coefficients to zero, selecting features. L2 adds squared weights and shrinks them smoothly.

k-fold cross-validation

The dataset is split into k folds; the model trains k times on k minus one folds and validates on the remaining fold, averaging the k validation scores.

Backpropagation

The algorithm that computes gradients of the loss with respect to each weight using the chain rule, working backward through the network layers.

Self-attention

A mechanism where each token forms query, key, and value vectors; scaled dot-product attention with softmax weights lets every token aggregate information from all others.

Reinforcement learning

An agent learns a policy that maximises cumulative reward by interacting with an environment, balancing exploration of new actions against exploitation of known good ones.

Overfitting

A model fits training noise rather than generalisable patterns, showing strong training performance and notably weaker validation or test performance.

NIST AI RMF functions

GOVERN establishes accountability and culture; MAP frames context and risks; MEASURE applies metrics and testing; MANAGE prioritises, responds to, and recovers from risks.

ISO/IEC 23894

A guidance standard on managing risk in AI systems, adapting the ISO 31000 process to AI-specific impacts across the system lifecycle.

Residual risk

The risk remaining after treatment controls are applied. It must be documented, assigned an owner, and formally accepted against the organisation's risk appetite.

Risk treatment options

Four choices for handling an identified risk: mitigate with controls, transfer to a third party such as insurance, avoid the activity, or accept it with justification.

Historical bias

Bias arising when training data reflects past inequities in the real world, so the model learns and reproduces those patterns even with perfect sampling.

Representation bias

Bias caused when the training sample under-represents certain groups or conditions, so model performance degrades for those populations at deployment.

Concept drift

A change over time in the statistical relationship between inputs and the target output, so a previously accurate model becomes stale and needs retraining.

Data drift

A shift in the distribution of input features compared with training data, which can degrade predictions even if the input-output relationship itself is unchanged.

Human in the loop

A design where a competent human reviews and can override AI outputs before consequential action, with real authority and sufficient time to intervene.

EU AI Act risk tiers

Four tiers: prohibited practices, high-risk systems subject to conformity assessment, limited-risk systems with transparency duties, and minimal-risk systems with no added obligations.

High-risk AI system

Under the EU AI Act, a system in Annex III areas such as employment, credit scoring, education, or critical infrastructure, subject to requirements on risk management, data governance, and oversight.

ISO/IEC 42001

A certifiable management system standard for AI, following Plan-Do-Check-Act and specifying requirements for leadership, risk assessment, operational controls, and continual improvement.

OECD AI Principles

Five intergovernmental principles: inclusive growth, human-centred values, transparency and explainability, robustness and safety, and accountability across the AI lifecycle.

AI system inventory

A register of deployed and planned AI use cases recording owner, model version, data sources, affected population, risk rating, oversight level, and next review date.

Transparency obligation

A duty to inform people that they are interacting with an AI system and to disclose meaningful information about its logic, outputs, and consequences.

AI governance roles

The governing body sets risk appetite; an AI risk owner approves use cases; model owners are accountable; independent review verifies controls.

Digital transformation stage gates

Decision checkpoints across ideation, pilot, scale, and retire phases that release funding only when predefined benefit and risk criteria are met.

Adversarial evasion attack

A deliberately perturbed input, often imperceptible to humans, crafted at inference time to make a trained model produce an incorrect output.

Data poisoning

An attack that corrupts or manipulates training data so the resulting model misbehaves on chosen inputs or carries an implanted backdoor trigger.

Model extraction

An attack that queries a deployed model repeatedly to reconstruct a functionally similar substitute, enabling further attacks and intellectual property loss.

Prompt injection

An attack where crafted text overrides an LLM's instructions. Direct injection comes from user input; indirect injection is hidden in retrieved documents or tool output.

Differential privacy

A guarantee that adding or removing one record changes outputs only marginally, controlled by a privacy budget epsilon; smaller epsilon means stronger privacy.

GDPR Article 22

Grants individuals rights over solely automated decisions with legal or similarly significant effects, including human intervention, expressing a view, and contesting the outcome.

OWASP LLM Top 10

A ranked list of large language model application risks, including prompt injection, insecure output handling, training-data poisoning, sensitive information disclosure, and excessive agency.