AWS Learning Material Certification Cheatsheets
AIF-C01 Domain 1 20% domain

Task 1.1: Explain basic AI concepts and terminologies

Fundamentals of AI and ML · 4,910 words · source: Certified-AI-Practitioner-AIF-C01/domain-1/task-1-1-cheat-sheet.md

Domain 1: Fundamentals of AI and ML

Task Statement 1.1: Explain basic AI concepts and terminologies

Domain 1 is 20% of scored exam content. Task 1.1 is the vocabulary foundation for the whole exam: AI vs ML vs deep learning vs GenAI vs agentic AI, data types, learning types, inference types, and basic quality concepts such as fit, bias, and fairness.

Use this sheet to answer definition, matching, ordering, and "which concept is described?" questions. For AIF-C01, you need exam-ready recognition more than math or implementation.

Exam Objective Map

You should be able to:

AWS exam objective What to know for the exam
Define basic AI terms AI, ML, deep learning, neural networks, computer vision, NLP, model, algorithm, training, inference, bias, fairness, fit, LLM, GenAI, agentic AI
Compare AI, ML, GenAI, deep learning, and agentic AI AI is the broad field; ML learns from data; deep learning uses neural networks; GenAI creates new content; agentic AI uses models plus tools/actions to pursue goals
Describe inference types Real-time, batch, asynchronous, serverless; choose by latency, traffic pattern, payload size, processing time, and cost
Describe AI/ML data types Labeled, unlabeled, structured, semi-structured, unstructured, tabular, time-series, image, text
Describe AI/ML learning types Supervised learning, unsupervised learning, reinforcement learning

The Big Picture

AI is the broadest term. Machine learning is a common way to build AI systems. Deep learning is a type of ML that uses neural networks. Generative AI is commonly powered by deep learning and foundation models. Agentic AI uses models, tools, memory, and orchestration to take actions toward goals.

Artificial intelligence
  Machine learning
    Traditional ML
    Deep learning
      Neural networks
      Foundation models
      Large language models
      Generative AI
  Agentic AI systems can use FMs, tools, APIs, memory, and orchestration

Exam shortcut:

AI = broad human-like capability.
ML = learns patterns from data.
Deep learning = ML with neural networks.
GenAI = creates new content.
LLM = language-focused foundation model.
Agentic AI = model plus tools/actions toward a goal.
Inference = model output on new input.

Core Terms

Term Exam-ready definition Example clue
Artificial intelligence (AI) Field of computer science focused on systems that perform tasks associated with human intelligence, such as learning, creation, reasoning, image recognition, or language understanding. "Human-like cognitive task"
Machine learning (ML) Branch of AI where systems learn patterns from data and improve predictions or decisions without explicit step-by-step rules. "Learns from historical data"
Deep learning Type of ML that uses neural networks with multiple layers to learn complex patterns, especially in unstructured data. "Images, speech, text, many layers"
Neural network Model structure made of connected nodes/layers that learn weights from data. "Input layer, hidden layers, output layer"
Computer vision AI that extracts information from images or video. "Detect scratches, faces, objects, labels"
Natural language processing (NLP) AI that understands, interprets, transforms, or generates human language. "Sentiment, entities, chatbot, translation, summarization"
Algorithm The mathematical procedure or method used to learn patterns or produce outputs. "Training adjusts parameters of an algorithm"
Model A trained artifact that uses learned patterns to produce predictions, classifications, recommendations, or generated outputs. "Deploy the model for inference"
Feature Input variable used by a model, such as a table column, pixel pattern, timestamp value, or text token. "Income, age, product price, image pixels"
Parameter / weight Internal value learned during training that controls model behavior. "Adjusted to reduce error"
Training Process of using data to adjust model parameters so the model learns patterns. "Model learns from examples"
Inference Using a trained model on new input to generate a prediction, score, classification, recommendation, or response. "Make a prediction on new data"
Prediction Probabilistic output from a model. It is an educated guess, not a guaranteed deterministic answer. "Probability of fraud"
Fit How well the model captures the relationship between inputs and outputs. "Underfit, good fit, overfit"
Bias Disparity or skew in model performance or outcomes across groups or classes. "Loan approvals worse for one group"
Fairness Goal of identifying, measuring, reducing, and monitoring unfair model behavior. "Evaluate outcomes across protected groups"
Generative AI (GenAI) AI that creates new content such as text, code, images, audio, video, summaries, or synthetic data. "Create, draft, generate, summarize"
Foundation model (FM) Large model trained on broad data that can be adapted to many tasks. "Reusable broad model"
Large language model (LLM) Foundation model focused on language understanding and generation. "Chat, summarize, write, answer, code"
Prompt Input instructions and context given to a generative model. "Ask Bedrock to summarize this article"
Response / completion Output generated by a generative model. "Generated answer or text"
Agentic AI AI system that can reason over a goal, use tools/APIs/data, and take actions through orchestration. "Open a ticket, call an API, complete a task"

Exam trap: inference is not the same as training. Training learns the model. Inference uses the trained model.

Similarities and Differences

Concept What it does Depends on Typical output
AI Broad umbrella for intelligent behavior Rules, ML, GenAI, search, planning, optimization Decisions, automation, predictions, generated content
ML Learns patterns from data Historical training data and algorithms Prediction, classification, score, recommendation
Deep learning Learns complex representations with neural networks Large datasets, more compute, layered networks Image labels, speech recognition, NLP outputs, GenAI outputs
GenAI Creates new content from prompts/context Foundation models, deep learning, prompts, retrieval, guardrails Text, code, image, audio, video, summary
Agentic AI Uses models plus tools/actions to pursue goals FM/LLM, instructions, tools, APIs, memory, orchestration Completed task, API action, workflow result

Memory hook:

Recognize or predict -> ML.
Complex unstructured patterns -> deep learning.
Create or transform content -> GenAI.
Plan and act with tools -> agentic AI.

AI, ML, Deep Learning, and GenAI

Question Traditional ML Deep learning Generative AI
Main strength Structured predictions and classifications Complex patterns in images, speech, text, and other unstructured data Creating or transforming content
Data fit Tabular, labeled, historical datasets Large image, video, audio, text, or mixed datasets Broad pretraining data plus prompts, retrieval, or customization
Human feature work Often needs feature engineering Learns features automatically from raw data better than traditional ML Uses learned broad representations in an FM
Compute need Often lower Often higher Often high, usually consumed through managed FM APIs such as Amazon Bedrock
Example Predict churn from customer records Detect defects in product images Draft email, summarize policy, generate code

Exam trap: GenAI is not just "better ML." It is best associated with generating new content and using foundation models. Traditional ML may still be better for structured prediction tasks.

Neural Networks

Deep learning models use neural networks with layers of nodes.

Component Meaning
Input layer Receives input features such as pixels, tokens, or tabular values
Hidden layers Learn intermediate patterns and representations
Output layer Produces the prediction, class, score, or generated output
Weight Learned value that controls how strongly a feature or node influences the result
Error / loss Difference between model output and expected output during training

Training adjusts weights to reduce error. In simple terms:

Data -> model prediction -> compare with expected result -> adjust weights -> repeat

Deep learning is strong for computer vision and NLP because it can learn complex relationships in pixels, words, audio, and other unstructured inputs.

Training vs Inference

Concept What happens Exam clue
Training Data is used to learn or adjust model parameters. "Build the model", "learn from examples", "adjust weights"
Model artifact Trained model files, learned parameters, model definition, and metadata. "Store artifacts in S3", "package for deployment"
Deployment Model artifacts and inference code are made available to applications. "Create endpoint", "host model", "run batch job"
Inference New input is sent to the trained model and output is returned. "Predict", "classify", "score", "generate response"

Exam shortcut:

Training consumes examples.
Inference consumes new input.
Model artifacts are the result of training.
Deployment makes inference possible.

Model Fit

Fit describes how well a model learned the useful relationship in data.

Fit state Meaning What happens Common fix
Underfitting Model is too simple or insufficiently trained. Poor performance on training data and new data. Improve features, use better model, train more, add useful data
Good fit Model captures useful pattern and generalizes. Performs well on training data and unseen data. Monitor and maintain
Overfitting Model memorizes training data or noise too closely. Performs well on training data but poorly on new data. Add diverse data, reduce complexity, regularize, stop training earlier

Exam trap:

Great training accuracy but poor new-data performance -> overfitting.
Poor training and new-data performance -> underfitting.

Bias and Fairness

Bias and fairness are responsible AI concepts that appear early in Domain 1 and return in Domain 4.

Concept Meaning Exam signal
Bias Model outcomes or performance are skewed for or against a group, class, or outcome. "Training data lacks diverse examples"
Fairness Models should be evaluated and designed to reduce unfair or discriminatory outcomes. "Check performance across demographic groups"
Data representation Training data should represent the population and scenarios where the model will be used. "No approved applications from a group in training data"
Feature risk Some features can introduce or proxy sensitive attributes. "Age, gender, location, protected class"
Continuous evaluation Fairness should be checked before deployment and after deployment. "Model behavior changes over time"

Exam shortcut:

Bad or incomplete data can create biased predictions.
Fairness starts before training and continues after deployment.

Data Types for AI Models

Data type Definition Examples Exam clue
Labeled data Training examples include the expected answer or target. Image labeled "fish"; transaction labeled "fraud" Supervised learning
Unlabeled data Training examples do not include known target answers. Customer behavior logs without segments Unsupervised learning
Structured data Organized in rows and columns with a defined schema. CSV, relational database table, Amazon RDS, Amazon Redshift SQL, tabular data
Semi-structured data Has organization but not strict rows/columns for every record. JSON, key-value records, DynamoDB, DocumentDB Flexible attributes
Unstructured data Does not fit a predefined tabular model. Images, video, audio, free-form text, social posts Object storage, embeddings, tokenization
Tabular data Structured data arranged like a table. Customer records, transactions, inventory Columns become features
Time-series data Records ordered by timestamp. Metrics, sensor readings, stock prices, sales by day Forecasting, trend detection
Image data Pixel-based visual data. Defect images, medical scans, faces Computer vision
Text data Natural-language content. Reviews, documents, chats, emails NLP, GenAI, embeddings

AWS storage cue:

Amazon S3 is a common storage location for ML training data and model artifacts because it can store many data types as objects.

Exam trap: "text" can be structured if it is a column in a table, semi-structured if it is in JSON, or unstructured if it is free-form documents or posts. Focus on how the data is organized.

Structured vs Semi-Structured vs Unstructured

Type How to recognize it Typical AWS examples
Structured Fixed schema, rows, columns, SQL-friendly Amazon RDS, Amazon Redshift, CSV files
Semi-structured Flexible fields or key-value structure JSON, Amazon DynamoDB, Amazon DocumentDB
Unstructured No fixed data model; content must be processed to extract features Amazon S3 objects, images, videos, audio, PDFs, text documents

Memory hook:

Rows and columns -> structured.
Keys and flexible attributes -> semi-structured.
Files, media, free text -> unstructured.

Learning Types

Learning type Training data Goal Output Example
Supervised learning Labeled data with known inputs and outputs Learn to predict the known target for new data Class, number, score Fraud/not fraud, house price, fish/not fish
Unsupervised learning Unlabeled data Find hidden patterns, groupings, anomalies, or structure Clusters, segments, anomaly signals Customer segments, network traffic groups
Reinforcement learning Agent interacts with an environment and receives rewards Learn a policy that maximizes long-term reward Actions or strategy AWS DeepRacer car learns to stay on track

Exam shortcut:

Known answers -> supervised.
No labels, discover patterns -> unsupervised.
Agent + environment + reward -> reinforcement learning.

Supervised Learning

Supervised learning trains with examples that include both input features and the desired output label or target.

Task Output type Example
Classification Category Fraud/not fraud, pass/fail, image category
Regression Numeric value Price, demand, delivery time
Recommendation Ranked item or score Product suggestions from historical behavior

Exam clues:

  • "Pre-labeled training data"
  • "Known answer"
  • "Predict a target"
  • "Classify into known categories"
  • "Predict a continuous number"

Common AWS anchor: Amazon SageMaker AI for building, training, and deploying custom ML models; SageMaker Ground Truth for data labeling workflows.

Unsupervised Learning

Unsupervised learning uses data with features but no known output labels.

Task What it does Example
Clustering Groups similar data points. Group customers by behavior
Anomaly detection Finds unusual patterns. Detect abnormal sensor values
Dimensionality reduction Simplifies data while preserving useful patterns. Prepare data for modeling or visualization
Topic discovery Finds themes in text. Group support tickets by theme

Exam clues:

  • "No labeled data"
  • "Find hidden patterns"
  • "Group similar records"
  • "Detect abnormal behavior"
  • "Discover categories automatically"

Exam trap: unsupervised learning and reinforcement learning can both work without labeled training data, but reinforcement learning has an agent, environment, actions, and rewards.

Reinforcement Learning

Reinforcement learning trains an agent through trial and error.

Term Meaning AWS Skill Builder example
Agent The learner or decision maker. AWS DeepRacer car
Environment The world the agent interacts with. Racetrack
State Current information about the environment. Car position and track state
Action Choice the agent takes. Turn, accelerate, brake
Reward Feedback signal for an action. Reward for staying on track and moving efficiently
Policy Strategy the agent learns to choose actions. Driving strategy

Use reinforcement learning when an autonomous agent must learn actions in a dynamic environment to maximize reward. It is not the default answer for ordinary classification, regression, or clustering.

Inference Types

SageMaker AI documents four core inference options. For AIF-C01, know when each one fits.

Inference type Best for How it works Exam clue
Real-time inference Low-latency online requests and sustained traffic Persistent managed endpoint handles requests immediately "User needs response now", "low latency", "high throughput"
Batch inference / batch transform Offline scoring when large data is available upfront Batch job runs, produces results, and shuts down compute "Monthly report", "score millions overnight", "no persistent endpoint"
Asynchronous inference Large payloads or long-running requests that can be queued Request is queued; result is returned later "Large payload", "processing can take minutes", "do not block client"
Serverless inference Intermittent or unpredictable traffic AWS manages infrastructure and scaling; pay for usage, not idle endpoint capacity "Spiky traffic", "avoid idle compute"

Decision shortcut:

Need immediate response with steady traffic -> real-time.
Have all records upfront and can wait -> batch.
Large/slow request and result can come later -> asynchronous.
Unpredictable intermittent requests -> serverless.

Exam trap: real-time is not always best. It can cost more when traffic is low because endpoint capacity is kept available.

Model Output Is Probabilistic

Most ML outputs are probabilities, scores, or generated responses rather than guaranteed truth.

Output Meaning Example
Probability Confidence-like estimate for a class or outcome. 0.93 probability the image is a fish
Score Numeric risk, relevance, or ranking value. Fraud risk score
Class Chosen category, often based on highest probability. "not fraud"
Generated response New content produced by a GenAI model. Summary, answer, code

Exam trap: if a scenario requires a specific deterministic result, Task 1.2 will usually push you away from ML and toward rules/application logic. For Task 1.1, remember that "inference" means a model output on new data.

Computer Vision and NLP

Concept What it handles Common tasks AWS service anchors
Computer vision Images and video Object detection, image classification, facial analysis, defect detection, OCR-like extraction Amazon Rekognition, Amazon SageMaker AI, Amazon Textract for document text extraction
NLP Human language text or speech-derived text Sentiment, entities, key phrases, classification, translation, summarization, chatbots Amazon Comprehend, Amazon Translate, Amazon Lex, Amazon Bedrock

Exam shortcut:

Pixels/video -> computer vision.
Words/language -> NLP.
New language content -> GenAI/LLM.

LLMs and Generative AI

Large language models are foundation models specialized for language. They are commonly built with transformer neural networks and trained on large text/code datasets.

Concept Know this
Prompt User instructions and context sent to the model
Response Generated output returned by the model
Transformer Neural network architecture strongly associated with modern LLMs
Foundation model Broad base model adaptable to many tasks
LLM FM focused on language tasks
Amazon Bedrock Fully managed AWS service for building and scaling GenAI applications with foundation models

Common LLM tasks:

  • Summarization
  • Question answering
  • Text generation
  • Code generation
  • Translation or rewriting
  • Classification through prompting
  • Conversational assistance

Exam trap: LLMs are not the same as all AI. They are one important type of foundation model focused on language.

Agentic AI

Agentic AI goes beyond one model response. It can use a model to reason about a goal, choose actions, use tools, and interact with APIs or data sources.

Component Meaning
Goal Task the agent is trying to complete
Model FM/LLM that interprets input and helps decide next steps
Tools APIs, functions, data sources, or applications the agent can call
Memory/context Information retained or retrieved to support the task
Orchestration Control logic that coordinates steps, tools, and responses

Exam clues:

  • "Take action"
  • "Use tools"
  • "Call APIs"
  • "Complete a multi-step workflow"
  • "Create a ticket, schedule an appointment, update a record"

Exam trap: a chatbot that only answers questions is not necessarily agentic. Agentic AI is signaled by planning and acting with tools or workflows.

Service Name Anchors for Task 1.1

Task 1.1 is mostly conceptual, but some AWS services are useful anchors.

Service Why it matters here
Amazon SageMaker AI Fully managed ML service for building, training, deploying, and hosting ML and foundation models; relevant for training, model artifacts, and inference options
Amazon S3 Common storage for training data, unstructured objects, and model artifacts
Amazon Bedrock Fully managed service for accessing foundation models and building GenAI applications
SageMaker Ground Truth Data labeling workflows for supervised learning datasets
AWS DeepRacer Learning example for reinforcement learning concepts
Amazon Rekognition Managed computer vision service anchor
Amazon Comprehend Managed NLP service anchor
Amazon Lex Conversational AI service anchor

Do not memorize implementation details here. Memorize what each service helps you recognize in a scenario.

Common Exam Traps

Trap Correct thinking
"AI, ML, and deep learning are synonyms." AI is broad; ML is a subset/approach; deep learning is a type of ML.
"Deep learning is always better." It can be powerful for unstructured data but often needs more data and compute.
"A model output is guaranteed correct." Inference is probabilistic and must be evaluated.
"High training performance proves the model is good." It might be overfitting. Check unseen data.
"Unlabeled data means reinforcement learning." Unlabeled pattern discovery is usually unsupervised learning. RL needs agent, environment, actions, rewards.
"Batch inference means slow real-time inference." Batch means offline processing of data available upfront.
"Serverless inference means no servers exist." AWS manages the infrastructure; it is best for intermittent/unpredictable traffic.
"Bias only comes from the algorithm." Bias can come from data, labels, features, design choices, and deployment context.
"Fairness is checked once." Fairness should be considered before training and monitored after deployment.
"All GenAI systems are agents." Agents use models plus tools/actions/orchestration; many GenAI apps only generate responses.

One-Page Memorization Version

  • AI is the broad field of systems that perform human-like intelligent tasks.
  • ML is AI that learns patterns from data instead of relying only on explicit instructions.
  • Deep learning is ML with neural networks and is strong for images, speech, text, and complex unstructured data.
  • Neural networks use input, hidden, and output layers; training adjusts weights to reduce error.
  • GenAI creates new content; foundation models are broad reusable models; LLMs are language-focused FMs.
  • Agentic AI uses a model plus tools, APIs, memory, and orchestration to take actions toward a goal.
  • Training creates a model; inference uses the trained model on new input.
  • Model artifacts include learned parameters, model definition, and metadata.
  • Inference outputs are usually probabilities, scores, classes, recommendations, or generated content.
  • Labeled data maps to supervised learning.
  • Unlabeled data maps to unsupervised learning.
  • Agent plus environment plus reward maps to reinforcement learning.
  • Structured data has rows and columns.
  • Semi-structured data has flexible key-value structure such as JSON.
  • Unstructured data includes images, video, audio, and free-form text.
  • Time-series data is timestamped and ordered.
  • Overfitting means good training performance but poor new-data performance.
  • Underfitting means poor performance on both training data and new data.
  • Bias means skewed model behavior across groups or outcomes.
  • Fairness means identifying and reducing unfair model outcomes.
  • Real-time inference is for low latency and sustained traffic.
  • Batch inference is for offline processing of large datasets available upfront.
  • Asynchronous inference is for queued, long-running, or large-payload requests.
  • Serverless inference is for intermittent or unpredictable traffic.
  • Amazon SageMaker AI anchors custom ML training, deployment, and inference options.
  • Amazon Bedrock anchors managed foundation models and GenAI applications.

Mini Practice Questions

  1. A company has images labeled "defective" and "not defective" and wants to train a model to classify new product images. Which learning type is this?

    • Answer: Supervised learning, because the training data includes labels.

  2. A team has unlabeled network traffic records and wants to discover unusual patterns. Which learning type fits?

    • Answer: Unsupervised learning, commonly clustering or anomaly detection.

  3. A simulated race car learns through rewards for staying on track and penalties for poor driving. Which learning type is this?

    • Answer: Reinforcement learning.

  4. A model performs extremely well on training data but poorly on new real-world examples. What is the issue?

    • Answer: Overfitting.

  5. A model cannot find a useful relationship and performs poorly on training and test data. What is the issue?

    • Answer: Underfitting.

  6. A company wants to process millions of customer records overnight and generate a report the next morning. Which inference type fits?

    • Answer: Batch inference / batch transform.

  7. A mobile app needs an immediate fraud score while a user is checking out. Which inference type fits?

    • Answer: Real-time inference.

  8. A model request can take several minutes and the client can retrieve the result later. Which inference type fits?

    • Answer: Asynchronous inference.

  9. A workload receives unpredictable bursts of requests and should not pay for idle endpoint capacity. Which inference type fits?

    • Answer: Serverless inference.

  10. A business user asks an AI system to summarize a long policy document and draft an email. Which AI concept is most directly involved?

    • Answer: Generative AI, likely using an LLM/foundation model.

  11. A system answers a customer question, calls an order API, and opens a return request. Which concept is this?

    • Answer: Agentic AI, because it uses tools/actions to complete a workflow.

  12. A dataset is a table of customers with columns for age, plan type, monthly spend, and churn status. What data type is it?

    • Answer: Structured/tabular data.

  13. A dataset contains JSON records where some items have fields that others do not. What data type is it?

    • Answer: Semi-structured data.

  14. A dataset contains product review paragraphs and support emails. What data type is it?

    • Answer: Unstructured text data.

  15. A loan model performs worse for a demographic group because the training data underrepresented approved applications from that group. What concept is involved?

    • Answer: Bias and fairness.

Quick Self-Check

  • I can explain AI, ML, deep learning, GenAI, LLMs, and agentic AI without treating them as synonyms.
  • I can distinguish training from inference.
  • I can identify model artifacts and basic deployment/inference ideas.
  • I can choose real-time, batch, asynchronous, or serverless inference from scenario clues.
  • I can distinguish labeled and unlabeled data.
  • I can distinguish structured, semi-structured, unstructured, tabular, time-series, image, and text data.
  • I can choose supervised, unsupervised, or reinforcement learning from scenario clues.
  • I can explain overfitting, underfitting, bias, and fairness at a foundational level.
  • I can map computer vision to images/video and NLP to human language.
  • I can recognize Amazon SageMaker AI, Amazon Bedrock, Amazon S3, SageMaker Ground Truth, and AWS DeepRacer as anchors for Task 1.1 concepts.

Sources

Official AWS and local project sources used: