AAIA flashcards

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

Supervised learning

A training approach that learns a mapping from inputs to outputs using labeled example pairs, enabling prediction on new data for classification or regression tasks.

Precision versus recall

Precision is the share of predicted positives that are correct; recall is the share of actual positives that were found. They trade off as the decision threshold changes.

Overfitting

A failure mode where a model memorizes training data, producing low training error but poor generalization to unseen data. Detected by a widening gap between training and validation metrics.

Bias-variance tradeoff

The tension between a model too simple to capture patterns (high bias, underfitting) and one too sensitive to training noise (high variance, overfitting). Model capacity controls this balance.

K-fold cross-validation

An evaluation method that partitions data into k folds, trains on k−1 folds, validates on the remaining fold, and averages results to reduce evaluation variance.

Data leakage

A modeling error where information unavailable at prediction time, such as target-derived features or validation data, contaminates training and inflates reported performance.

Self-attention mechanism

A transformer operation computing softmax(QKᵀ/√d_k)V, which lets each token weigh every other token in the context window when building its representation.

Temperature in sampling

A parameter scaling logits before softmax. Lower values sharpen the probability distribution for deterministic output; higher values increase diversity and hallucination risk.

Context window

The maximum number of tokens a language model can process in one request, bounding the combined length of system prompt, user input, retrieved context, and output.

Hallucination

Fluent model output that is unsupported or factually wrong. It arises because training optimizes token likelihood rather than truth, and is reduced by grounding and citation checks.

Retrieval-augmented generation (RAG)

A pattern that embeds a query, retrieves relevant document chunks from a vector store, and inserts them into the prompt so the model answers from supplied sources.

Tokenization

The preprocessing step splitting text into subword units called tokens, which the model embeds and predicts. Token counts drive context limits, latency, and cost.

Top-p nucleus sampling

A decoding strategy keeping the smallest set of tokens whose cumulative probability exceeds p, then sampling among them to balance coherence and variety.

Diffusion model

A generative model trained to reverse a noising process, producing images or audio by iteratively denoising random input toward a coherent sample.

Few-shot prompting

A prompting technique placing several input-output examples in the prompt so the model infers the desired task format without weight updates.

RAG versus fine-tuning

Choose RAG for changing or proprietary knowledge needing citations; choose fine-tuning for fixed style, format, or behavior. They can be combined in one application.

Vector database

A store indexing dense embeddings and answering nearest-neighbor queries, often approximately via HNSW or IVF, to supply relevant context for retrieval-augmented generation.

Function calling

A model capability to emit structured arguments matching a declared tool schema, letting application code execute APIs and return results to the model.

LLM-as-judge evaluation

Using a language model to score outputs against criteria at scale. It inherits judge bias and position effects, so it must be calibrated against human labels.

Multimodal model

A foundation model accepting more than one input modality, such as text plus images or audio, enabling tasks like visual question answering and document extraction.

Chunking strategy

The rule splitting source documents into retrieval units. Chunk size and overlap strongly affect retrieval precision and the context passed to the model.

Demographic parity

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

Equalized odds

A fairness criterion requiring equal true positive and false positive rates across groups, so error patterns do not differ by protected attribute.

SHAP explanations

A method attributing a prediction to features using Shapley values from cooperative game theory, giving consistent local and global importance measures.

Model card

A transparency document stating a model's intended use, limitations, evaluation metrics, and ethical considerations so downstream users can assess suitability.

Differential privacy

A guarantee adding calibrated noise so query results reveal little about any individual record, quantified by an epsilon privacy budget.

Automation bias

The tendency of human reviewers to over-trust automated recommendations, reducing the effectiveness of human oversight unless uncertainty is surfaced.

Red-teaming

Adversarial testing that probes a system for harmful, biased, or unsafe outputs before release, with findings feeding mitigation and documented risk acceptance.

Prompt injection

An attack where untrusted content, such as a retrieved document, contains instructions that override the developer's prompt and redirect model behavior.

NIST AI RMF functions

The framework's four core functions: Govern, Map, Measure, and Manage. They structure identification, assessment, and treatment of AI risks across the lifecycle.

EU AI Act risk tiers

A regulation classifying AI by risk: prohibited practices, high-risk systems with strict obligations, limited-risk transparency duties, and minimal-risk systems.

Data poisoning

An attack corrupting training data to implant backdoors or degrade model behavior, mitigated by provenance checks, data validation, and anomaly monitoring.