AWS Machine Learning and Generative AI Vocabulary Review

Flashcards covering AWS AI/ML services, generative AI concepts, model evaluation metrics, and security governance terminology based on the lecture keywords list.

Amazon SageMaker

A fully managed service that data scientists and developers use to prepare, build, train, and deploy ML models.

Amazon Bedrock

A fully managed service that makes FMs from Amazon and leading AI companies available through an API, allowing for private customization with your own data.

Knowledge Base

The specific implementation for Retrieval Augmented Generation (RAG) within Amazon Bedrock.

SageMaker Data Wrangler

A data preparation and feature engineering tool that supports SQL and offers balancing operators like random undersampling, random oversampling, and SMOTE (Synthetic Minority Oversampling Technique).

Amazon SageMaker Model Cards

A tool used to record information about ML models for governance and reporting, including training details, risk ratings, and evaluation metrics.

SageMaker Canvas

A service used to build ML models without writing any code, though it cannot perform content moderation for creative content types.

Amazon SageMaker Ground Truth

A service that uses a human workforce, including Amazon Mechanical Turk, to create accurate labels for training data.

Amazon SageMaker Model Monitor

An automated alert system that detects variations in model quality, such as data drift, anomalies, and performance issues.

SageMaker Studio

An integrated development environment (IDE) suite that includes JupyterLab, RStudio, and Visual Studio Code - Open Source (Code-OSS).

Guardrails for Amazon Bedrock

A safeguard layer that evaluates user inputs and FM responses to block undesirable content, prevent hallucinations, and redact PII.

Amazon Rekognition

A fully managed AI service for image and video analysis used for identifying inappropriate content and performing content moderation.

Vector Database

A collection of data stored as mathematical representations (vector embeddings) consisting of numbers that represent words, phrases, or images.

Amazon DocumentDB

A fully managed, native JSON document database that supports vector search with millisecond response times.

Amazon OpenSearch Service

A fully managed service used for semantic search, RAG, and recommendation engines, capable of scaling to millions of embeddings.

Amazon SageMaker Clarify

A service that identifies potential bias in datasets and models and explains how a model makes predictions using feature attributions like Shapley values.

SageMaker JumpStart

A machine learning hub within SageMaker containing open-source and proprietary foundation models.

SageMaker HyperPod

A service that reduces foundation model training time by up to $40\%$ by efficiently parallelizing workloads across more than a thousand AI accelerators.

SageMaker Model Registry

A fully managed catalog used to manage model versions, metadata, and approval statuses.

Amazon Q Business

A generative AI virtual assistant that handles tasks based on provided data but does not provide access to foundation models (FMs).

Retrieval Augmented Generation (RAG)

The process of optimizing LLM output by referencing an authoritative, external knowledge base outside of its training data before generating a response.

Fine-Tuning

The process of further training a pre-trained model on a specific task using a labeled dataset to improve accuracy and reduce bias.

RLHF

Reinforcement Learning from Human Feedback; a technique for aligning LLMs with human values by rewarding behaviors based on human feedback.

Chain of Thought

A prompt engineering technique that breaks down complex questions into smaller parts, recommended for arithmetic and logical tasks.

Overfitting

When a model performs well on training data but fails to generalize to new data; characterized by low bias and high variance.

Underfitting

When a model is too simple to capture underlying patterns, characterized by high bias and low variance.

Explainability

The ability to understand and explain in human terms why a model arrived at a specific prediction (The "Why").

Interpretability

The ability to understand the internal decision-making process of a model and how it combines features (The "What").

ROUGE

An evaluation metric for the quality of generated summaries or translations compared to reference texts.

BLEU

An evaluation metric used to measure the similarity between generated text translations and reference translations.

Top K

A parameter that limits the selection of tokens to the $K$ most probable options during text generation.

Top P

A setting on a scale of $0-1$ that controls diversity by limiting word choice to a specific percentage of the total probability distribution (e.g., $0.25$ for the top $25\%$).

Temperature

A setting that controls randomness; lower values lead to deterministic responses, while higher values increase diversity and unpredictability.

F1 Score

A classification metric calculated as $F1 = \frac{2 \times P \times R}{P + R}$, where $P$ is Precision and $R$ is Recall.

Mean Squared Error (MSE)

The average of the squared differences between predicted and actual values; used for regression models where smaller values indicate better fit.

Root Mean Squared Error (RMSE)

The square root of the squared difference between predicted and actual values, used to indicate large model errors and outliers.

R squared ($R^2$)

A measurement from $0$ to $1$ indicating how much variation in data is explained by the model (e.g., $0.8$ means $80\%$ variation explained).

Hallucination

Incorrect or misleading results generated by AI models caused by insufficient data, incorrect assumptions, or biases.

GAN (Generative Adversarial Network)

A deep learning architecture where two neural networks compete to generate authentic new data.

Transformers

A neural network architecture that transforms input sequences into output sequences by learning context and tracking relationships.

SageMaker Asynchronous Inference

An inference type that queues incoming large requests (less than $1\,\text{GB}$) and processes them asynchronously.

Amazon Augmented AI (Amazon A2I)

A service used to build human review systems for ML solution predictions to audit individual results.

Epoch

A single complete pass through the entire training dataset during the training process.

Learning Rate

A hyperparameter that determines the step size at which a model's parameters are updated during training.

Context Window

A model property describing the number of tokens the model can accept as input.

Poisoning

The intentional introduction of malicious or biased data into a training dataset to cause harmful outputs.

Jailbreaking

The practice of using carefully constructed prompts to bypass safety measures and constraints implemented in a generative model.

Amazon Polly

A service that converts text to high-quality, natural-sounding speech in multiple languages with low operational overhead.

Amazon Lex

A fully managed AI service for building conversational interfaces like chatbots using voice and text.

Amazon Kendra

An intelligent document search service that uses semantic and contextual understanding to provide specific answers from data.

Amazon Macie

A security service that uses ML to discover and protect sensitive data (PII) stored in Amazon S3.

Amazon Inspector

An automated security assessment service that scans EC2 applications, Lambda code, and ECR repositories for vulnerabilities.