AIF flashcards

32 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, enabling the model to predict labels or values for new, unseen inputs.

Unsupervised learning

A machine learning approach that finds patterns, groups, or structure in unlabeled data without predefined target outputs or correctness labels.

Classification vs regression

Classification predicts discrete categories such as spam or not spam, while regression predicts continuous numeric values such as price, temperature, or demand.

Precision

The fraction of positive predictions that are actually positive, calculated as true positives divided by true positives plus false positives.

Recall

The fraction of actual positives that are correctly identified, calculated as true positives divided by true positives plus false negatives.

Overfitting

A model fits training data too closely, including noise, so training error is low but validation error is high and generalization is poor.

Bias-variance tradeoff

The balance between a model too simple to capture patterns, causing high bias, and a model too complex, causing high variance and overfitting.

Foundation model

A large model pretrained on broad data that can be adapted through prompting, retrieval, or fine-tuning to many downstream tasks.

Token

A subword or character unit produced by a tokenizer; models process and bill text as tokens, and context limits count input plus output tokens.

Context window

The maximum number of tokens a model can process in one request, including system prompt, user input, retrieved content, and generated output.

Temperature

A sampling parameter that controls randomness; lower values make output more deterministic and higher values increase diversity and creativity.

Embedding

A dense numeric vector representing text, images, or other content so semantic similarity can be measured by distance or cosine score.

Retrieval-augmented generation

A technique that retrieves relevant documents from a knowledge base and inserts them into the prompt so the model grounds answers in current external content.

Hallucination

A confident but unsupported or incorrect model output; mitigation includes retrieval grounding, citations, evaluation, guardrails, and human review.

Amazon Bedrock

A managed AWS service that provides API access to multiple foundation models, plus features for guardrails, knowledge bases, agents, and model evaluation.

Few-shot prompting

A prompting pattern that includes several example input-output pairs in the prompt so the model can infer the desired task and response format.

Chain-of-thought prompting

A prompting technique that asks the model to show intermediate reasoning steps, often improving performance on multi-step math, logic, and planning tasks.

Top-k sampling

A decoding method that restricts each next-token choice to the k most probable tokens, reducing unlikely outputs while preserving some randomness.

Model distillation

Training a smaller student model to imitate a larger teacher model's outputs, reducing inference cost and latency while retaining much of the performance.

Knowledge base

A repository of documents that are chunked, embedded, and indexed so a retrieval system can supply relevant context to a generative model.

Responsible AI

A set of design and governance principles ensuring AI systems are fair, explainable, privacy-preserving, safe, transparent, accountable, and controllable throughout their lifecycle.

Bias

Systematic unfairness in model outputs caused by skewed data, sampling, labeling, or proxy features, often measured across demographic subgroups.

SHAP

A model explanation method that assigns each feature a contribution value based on cooperative game theory, helping explain individual predictions.

Human-in-the-loop

A control pattern where a person reviews, approves, or overrides AI outputs before consequential actions, reducing harm and automation bias.

Bedrock Guardrails

A managed AWS feature that filters harmful content, denies topics, blocks sensitive information, and applies safety policies to foundation model inputs and outputs.

Demographic parity

A fairness metric requiring positive prediction rates to be similar across groups; it can conflict with accuracy and other fairness definitions.

Shared responsibility model

AWS secures the cloud infrastructure, while customers secure their data, identities, configurations, prompts, and applications running in the cloud.

Least privilege

An access principle granting only the permissions needed for a task, reducing blast radius if credentials or workloads are compromised.

Prompt injection

An attack where malicious input manipulates a model to ignore instructions, leak data, or invoke tools improperly; defenses include validation and guardrails.

Data poisoning

An attack that corrupts training or retrieval data to bias or degrade model behavior; controls include source validation, lineage, and access restrictions.

AWS CloudTrail

An AWS service that records API activity and account actions, supporting auditing, incident investigation, and compliance evidence for AI workloads.

Model drift

Degradation of model performance when real-world data changes from training conditions; monitoring, retraining, and evaluation thresholds detect and mitigate it.