HIRAF flashcards

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

Turing test

A behavioral test of machine intelligence in which a human judge converses with an unseen machine and human; the machine passes if the judge cannot reliably tell them apart.

Supervised learning

A machine learning paradigm that trains a model on labeled input-output pairs so it can predict outputs for new, previously unseen inputs.

Reinforcement learning

A learning paradigm where an agent takes actions in an environment and improves its policy by maximizing cumulative reward received over time.

Overfitting

A modeling failure where a model captures noise specific to training data, producing low training error but poor performance on new, unseen data.

Narrow AI

An AI system designed for a specific bounded task, such as translation or recommendation; all currently deployed production AI systems are narrow.

Bias-variance tradeoff

The tension where simpler models underfit with high bias, complex models overfit with high variance, and total error combines both plus irreducible noise.

Cross-validation

A resampling technique that splits data into k folds, training on k-1 and validating on the remaining fold, rotating until every fold is used.

Precision

A classification metric equal to true positives divided by the sum of true positives and false positives; it measures how often positive predictions are correct.

Recall

A classification metric equal to true positives divided by the sum of true positives and false negatives; it measures coverage of actual positive cases.

Data leakage

Contamination of training features with information unavailable at prediction time or from the test set, which inflates validation scores and misleads model selection.

L2 regularization

A penalty proportional to the squared magnitude of model weights, added to the loss, which shrinks coefficients and reduces variance without forcing them to zero.

Hyperparameter

A configuration value set before training begins, such as learning rate, tree depth, or regularization strength, that controls how the learning algorithm behaves.

Data lineage

Metadata that records the origin, movement, and transformation history of data, enabling traceability, impact analysis, and reproducibility of derived datasets.

Data minimization

A governance principle requiring collection and retention of only the personal data necessary for a specified, legitimate purpose, reducing privacy and compliance exposure.

Data steward

A role responsible for day-to-day data quality, definitions, and standards within a domain, operating under the accountability of the data owner.

Metadata

Structured information describing data, such as schema, source, owner, sensitivity classification, and refresh schedule, used to manage and locate datasets.

Data classification

Assignment of data to sensitivity tiers such as public, internal, confidential, or restricted, which determines required access, encryption, and retention controls.

Demographic parity

A fairness criterion requiring equal positive prediction rates across protected groups, regardless of differences in underlying base rates or outcome prevalence.

Proxy variable

An input feature correlated with a protected attribute, such as postal code or name, that can reproduce discriminatory outcomes even when the attribute itself is excluded.

Post hoc explanation

An interpretation method applied after training, such as SHAP or LIME, that approximates why a specific model produced a particular prediction.

Human-in-the-loop

An oversight design in which a human reviews and approves or overrides a system's decision before it takes effect, used for higher-risk or irreversible actions.

Residual risk

The risk remaining after controls and mitigations are applied; it must be documented, accepted by an accountable owner, and periodically reviewed.

Concept drift

A change over time in the statistical relationship between model inputs and the target outcome, degrading a deployed model's predictive accuracy.

EU AI Act risk tiers

A regulatory classification of AI systems into prohibited practices, high-risk, limited-risk with transparency duties, and minimal-risk, with obligations scaling by tier.

GDPR purpose limitation

A data protection principle requiring personal data to be collected for specified, explicit, legitimate purposes and not further processed incompatibly with them.

NIST AI RMF

A voluntary risk management framework organized around four functions, Govern, Map, Measure, and Manage, for identifying and treating AI risks across the lifecycle.

Model card

A short standardized document describing a model's intended use, training data, evaluation metrics, limitations, and ethical considerations for downstream users.

Model drift

Degradation of a deployed model's performance as live data diverges from training conditions; monitored via input distributions, prediction distributions, and outcome metrics.

Canary release

A deployment pattern that routes a small fraction of live traffic to a new model version, monitoring metrics before gradually expanding exposure or rolling back.

Feature store

A centralized system that computes, stores, and serves features consistently for both training and inference, reducing training-serving skew and duplicated pipeline work.

Shadow deployment

A deployment pattern in which a candidate model receives live inputs and its outputs are logged but not served, allowing comparison against production without user impact.

Training-serving skew

A mismatch between how features are computed during model training and during live inference, producing predictions that differ from validated offline performance.