AIAI flashcards

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

Supervised learning

A machine learning approach that trains on labeled input-output pairs so the model learns a mapping from features to a target, enabling prediction on new inputs.

Reinforcement learning

A learning paradigm in which an agent takes actions in an environment and learns a policy that maximizes cumulative reward through trial and error.

Overfitting

A modeling failure where the model fits training data noise too closely, producing low training error but poor performance on unseen data.

Hyperparameter

A configuration value set before training that governs the learning process, such as learning rate, batch size, or tree depth, and is not learned from data.

Narrow AI

An AI system designed for a specific, bounded task, such as translation or image classification; all deployed AI systems today are narrow.

Precision and recall

Precision is true positives divided by predicted positives; recall is true positives divided by actual positives. They trade off against each other.

Cross-validation

A resampling technique that partitions data into k folds, trains on k−1 and validates on the remainder, rotating folds to average a more stable performance estimate.

Concept drift

A change over time in the relationship between input features and the target variable, which degrades model accuracy and triggers retraining.

Confusion matrix

A table of predicted versus actual classes showing true positives, true negatives, false positives, and false negatives, from which precision and recall are computed.

Regularization

A penalty added to the loss function that discourages large model weights, reducing overfitting; L1 encourages sparsity and L2 shrinks weights smoothly.

Data leakage

Contamination of training data with information about the target or test set, producing unrealistically high validation scores that fail in production.

One-hot encoding

A categorical encoding that creates one binary column per category, avoiding any implied ordering; it increases dimensionality as category count grows.

Feature scaling

Transforming numeric features to a common range (normalization) or to zero mean and unit variance (standardization), required by distance-based algorithms.

Class imbalance

A condition where one class vastly outnumbers another, making accuracy misleading; remedies include resampling, class weights, and threshold adjustment.

Data completeness

A data quality dimension measuring the proportion of required values that are present and non-null; gaps require imputation or exclusion before modeling.

Self-attention

A transformer mechanism in which each token computes weighted relevance to all other tokens in the context, capturing long-range dependencies in parallel.

Retrieval-augmented generation

A technique that retrieves relevant documents and inserts them into the prompt so the model grounds its answer in external sources without weight updates.

Hallucination

Fluent model output that is factually incorrect or unsupported by any source, arising because the model optimizes token likelihood rather than factual truth.

Context window

The maximum number of tokens a language model can process in a single interaction, spanning both the input prompt and the generated output.

Temperature

A sampling parameter that scales output probabilities; lower values make generation more deterministic and higher values increase diversity and randomness.

Tokenization

The preprocessing step that splits text into subword units called tokens, which are converted to numeric IDs before being processed by a language model.

Demographic parity

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

Explainability

The degree to which a human can understand why a specific model produced a particular output, supported by methods such as SHAP and LIME.

Human-in-the-loop

A design pattern where a person reviews or authorizes consequential AI outputs, with authority and context sufficient to override the system.

Model card

A document describing a model's intended use, training data, evaluation results across subgroups, and known limitations to support responsible deployment.

Prompt injection

An attack in which malicious text in user input or retrieved content overrides system instructions, causing the model to follow attacker-controlled directions.

Differential privacy

A guarantee that adds calibrated noise so query results reveal little about any single individual, bounding the privacy loss of published statistics.

Data poisoning

An attack that corrupts training data to degrade model performance or implant backdoor behavior triggered by specific input patterns.

Federated learning

A training approach where models are trained locally on distributed devices and only parameter updates are shared, keeping raw data on the source.

Anomaly detection

A use case that identifies rare observations deviating from expected patterns, applied to fraud, network intrusion, and equipment fault detection.

Build versus buy

A decision comparing in-house development against purchasing a vendor solution, weighing differentiation, cost, time to market, and data control.

Canary release

A deployment strategy that routes a small fraction of live traffic to a new model version, limiting blast radius before full rollout.

Model registry

A versioned repository storing trained model artifacts with metadata and lifecycle stage, enabling reproducible deployment, rollback, and audit.