AWS Certified AI Practitioner (AIF-C01) – Key Vocabulary

Vocabulary flashcards covering core AWS AI services, machine-learning concepts, responsible-AI principles, and key metrics for the AWS Certified AI Practitioner exam.

Accuracy

Metric that measures the proportion of correct predictions out of all predictions; can be misleading on imbalanced datasets.

Activation Function

Mathematical function in a neural-network node that introduces non-linearity (e.g., ReLU, Sigmoid, Tanh).

Agents for Amazon Bedrock

Bedrock feature that lets foundation models plan and execute multi-step tasks through API calls across company systems.

Amazon Bedrock

Fully managed AWS service providing API access to multiple high-performing foundation models for building generative-AI apps.

Amazon Comprehend

NLP service that extracts entities, key phrases, sentiment, language, and more from unstructured text.

Amazon Kendra

Enterprise search service that answers natural-language questions by searching across multiple content repositories.

Amazon Lex

Service for building chatbots and voice bots using automatic speech recognition and natural-language understanding.

Amazon Polly

Text-to-speech (TTS) service that converts written text into lifelike spoken audio in multiple voices and languages.

Amazon Rekognition

Computer-vision service for image and video analysis (object detection, facial analysis, content moderation, etc.).

Amazon SageMaker

End-to-end managed platform to build, train, and deploy machine-learning models at scale.

Amazon Titan

Family of AWS-built foundation models (text, image, embeddings) available exclusively through Amazon Bedrock.

Amazon Transcribe

Automatic speech-recognition (ASR) service that turns spoken audio into text.

Amazon Translate

Neural machine-translation service for fast, high-quality translation between languages.

Artificial Intelligence (AI)

Broad field of computer science aimed at creating systems that perform tasks requiring human intelligence.

Bias (AI)

Systematic error causing prejudiced outcomes, often due to non-representative or biased training data.

Chatbot

Software application that simulates human conversation via text or voice.

Classification

Supervised-learning task that predicts a categorical label (e.g., spam vs. not spam).

Clustering

Unsupervised-learning technique that groups similar data points without pre-existing labels.

Computer Vision

AI field enabling computers to interpret and derive information from images or videos.

Context Window

Maximum number of tokens a foundation model considers in one request (prompt + output).

Deep Learning

Subset of machine learning that uses multilayer neural networks to learn complex patterns.

Domain-Adaptation Fine-Tuning

Technique that adapts a foundation model to a specific domain by training on large unlabeled domain text.

Embeddings

Vector representations that capture semantic meaning of text, images, or other data for tasks like search.

Epoch

One complete pass through the entire training dataset during model training.

Explainability

Responsible-AI principle focused on understanding and communicating how a model makes its decisions.

Fairness

Responsible-AI pillar ensuring model predictions are not biased against protected subgroups.

Fine-Tuning

Further training a pre-trained foundation model on a smaller labeled dataset to specialize it.

Foundation Model (FM)

Large, pre-trained deep-learning model that can be adapted for many downstream tasks (e.g., GPT-4, Titan).

Generative AI

AI class that creates new content (text, images, code, etc.) resembling the data it was trained on.

Hallucination

When a generative model outputs false or nonsensical information while appearing confident.

Hyperparameter

User-set configuration (e.g., learning rate, batch size) that governs the training process but is not learned.

Inference

Using a trained model to generate predictions or decisions on new, unseen data.

Instruction-Based Fine-Tuning

Supervised fine-tuning method where training data is formatted as prompt/completion instruction pairs.

Knowledge Bases for Amazon Bedrock

Managed feature that implements Retrieval-Augmented Generation by linking Bedrock models to company data in S3.

Labeled Data

Dataset where each sample is tagged with the correct output; required for supervised learning.

Large Language Model (LLM)

Type of foundation model trained on massive text corpora to understand and generate human-like language.

Latency

Time delay between sending a prompt and receiving a model’s response; critical for real-time apps.

Machine Learning (ML)

AI subset where algorithms learn patterns from data to make predictions without explicit programming.

Modality

Type of data a model processes—text, images, audio, video; a model handling multiple types is multimodal.

Multi-Turn Messaging

Stateful dialogue with multiple back-and-forth exchanges, limited by the model’s context window.

Natural Language Processing (NLP)

AI field focused on enabling computers to understand, interpret, and generate human language.

Neural Network

Computing architecture inspired by the brain, consisting of interconnected layers of artificial neurons.

Overfitting

When a model learns noise in training data, harming its performance on new data.

Parameter

Internal value (weight or bias) learned by a model during training.

Precision

Percentage of positive predictions that are actually correct; True Positives ÷ (True Positives + False Positives).

Prompt

Input text provided to a foundation model instructing it to perform a task or generate content.

Prompt Engineering

Crafting and refining prompts to steer a foundation model toward desired outputs.

Provisioned Throughput

Bedrock feature that reserves dedicated inference capacity for consistent, scalable performance.

Recall

Percentage of actual positives correctly identified; True Positives ÷ (True Positives + False Negatives).

Regression

Supervised-learning task that predicts a continuous numerical value.

Reinforcement Learning

ML paradigm where an agent learns to make sequential decisions by maximizing cumulative reward.

Responsible AI

Designing, developing, and deploying AI that is safe, ethical, fair, and transparent.

Retrieval-Augmented Generation (RAG)

Framework that improves LLM outputs by retrieving relevant knowledge and appending it to the prompt.

Single-Turn Messaging

Stateless interaction consisting of one user input and one model output with no memory of past turns.

Supervised Learning

Machine-learning approach where models learn from labeled data; includes classification and regression.

Token

Basic unit of text (word, sub-word, or punctuation) processed by a language model.

Training

Process of feeding data to a model so it learns the values of its parameters for accurate predictions.

Transparency

Responsible-AI principle that provides clear information on an AI system’s purpose, workings, and limitations.

Unlabeled Data

Dataset without tags or annotations; used in unsupervised learning.

Unsupervised Learning

Machine-learning approach that discovers patterns in unlabeled data (e.g., clustering).

Validation Data

Subset of data used during training to evaluate model performance and tune hyperparameters, helping prevent overfitting.

EC2 Bootstrapping

is the process of running scripts on an EC2 instance when it first launches to automatically install software and apply configurations. This allows for immediate setup and customization of the instance without manual intervention.