AIF-C01 Concept Map (Mind Map): Domain 1

Every topic from Domain 1 of the exam on one page, with the pairs the exam swaps to trick you. Domain 1 is 20% of your score.

1.1 Concepts and terminology

The vocabulary every other domain builds on

Artificial intelligence: Any technique that lets machines do tasks that normally need human intelligence.
Machine learning: Systems that learn patterns from data instead of following rules written by people.
Deep learning: ML using neural networks with many layers. Strongest on unstructured data like images, audio and text.
Neural network: Layers of connected nodes whose weights adjust during training.
Generative AI: Deep learning models that create new content: text, images, code, audio.
Large language model: A foundation model trained on huge amounts of text to understand and generate language.
Agentic AI: AI that plans multi-step tasks, uses tools and takes actions toward a goal with limited human input.

Algorithm: The learning method, such as linear regression or a decision tree.
Model: The trained artifact that makes predictions on new data.
Training: Feeding data to an algorithm so it learns patterns. Heavy, done occasionally.
Inference: Using a trained model to produce predictions or outputs on new input.
Fit: over and under: How well a model generalizes. Overfit: memorized the training data, fails on new data. Underfit: too simple, poor everywhere.
Bias: A systematic skew in data or predictions that disadvantages some groups or outcomes.
Fairness: Model outcomes that are equitable across groups.
Computer vision: AI that interprets images and video.
Natural language processing: AI that understands and processes human language.

Real-time: A persistent endpoint that answers each request with low latency.
Batch: Predictions on a whole dataset at once, offline, with no persistent endpoint (Batch Transform).
Asynchronous: Requests are queued and processed in the background. Built for large payloads and long processing times; can scale to zero.
Serverless: An endpoint that scales automatically, down to zero. Pay per use; cold starts are possible.

Labeled: Data with the correct answer attached, like an email marked “spam”.
Unlabeled: Raw data with no answers attached.
Structured: Organized in rows and columns: tabular data and time-series.
Unstructured: No fixed schema: text, images, audio, video.

Supervised: Learns from labeled examples to predict labels. Covers classification and regression.
Unsupervised: Finds hidden structure in unlabeled data, such as clusters or anomalies.
Reinforcement: An agent learns by trial and error, maximizing rewards from its environment.

Matching the problem to the technique and the service

When AI adds value: Supporting human decisions, scaling beyond human capacity, automating repetitive work, spotting patterns too complex for rules.
When AI is the wrong tool: When cost outweighs benefit, or when you need an exact, guaranteed outcome instead of a prediction.
Traditional ML: Narrow task, structured data, explainable, cheap and fast to run.
Foundation model: Large pre-trained model for broad tasks: generating, summarizing, reasoning over unstructured content.

Fraud detection: Classification or anomaly detection on transactions.
Forecasting: Predicting future values from time-series data: a regression problem.
Recommendations: Suggesting items based on behavior and similarity.
Knowledge bases: Letting a foundation model answer from your own documents (RAG). On AWS: Amazon Bedrock Knowledge Bases.
Speech recognition: Turning spoken audio into text.

SageMaker AI: Build, train and deploy your own ML models end to end.
Amazon Transcribe: Speech → text.
Amazon Polly: Text → lifelike speech.
Amazon Translate: Text in one language → text in another.
Amazon Comprehend: NLP insights from text: sentiment, entities, key phrases, language, PII.
Amazon Lex: Build chatbots and voice bots with intents and slots; the technology behind Alexa.
Amazon Rekognition: Image and video analysis: objects, faces, text in images, content moderation.
Amazon Textract: Extracts text, forms and tables from scanned documents.
Amazon Personalize: Real-time personalized recommendations, no ML expertise needed.
Amazon Kendra: Intelligent enterprise search over your documents in natural language. Closed to new customers since July 30, 2026: new projects use Amazon Bedrock Knowledge Bases, but practice questions still name Kendra.

From raw data to a monitored model in production

Collect data: Gather data and label it if needed.
Prepare and explore: Clean, transform and explore the data (EDA).
Engineer features: Create and select the input variables the model learns from.
Train: Fit the algorithm to the data.
Tune: Search for the best hyperparameters.
Evaluate: Test on held-out data; check metrics, bias and explainability.
Deploy: Put the model into production; version and approve it in the registry.
Monitor: Watch data and prediction quality in production, and trigger retraining.

Open-source pre-trained: Start from an existing model and adapt it, for example via SageMaker JumpStart or Hugging Face.
Custom, from scratch: Train your own model: full control, but needs lots of data, time and money.

Managed API: Call a model through an API, as with Amazon Bedrock; AWS runs the infrastructure.
Self-hosted API: You deploy and operate the model yourself, on a SageMaker endpoint or EC2.

Amazon Bedrock: Fully managed access to foundation models from Amazon and other providers, plus knowledge bases, agents and guardrails.
Amazon Quick: AWS’s AI workspace for business users: agents, research, automation and BI dashboards over company data.
Kiro: AWS’s agentic development environment: AI agents turn specs into working code.

MLOps practices: Experiment tracking, repeatable pipelines, scalable systems, managing technical debt, production readiness.
Monitoring and retraining: Real-world data changes and accuracy decays, so you detect drift and retrain.

Accuracy: Share of all predictions that are correct.
Precision: When the model says yes, how often is it right?
Recall: Of all the real positives, how many did the model find?
F1 score: Balances precision and recall in one number (their harmonic mean).
Business metrics: Cost per user, development cost, customer feedback, return on investment.

Part of the AWS Certified AI Practitioner Study Kit: Concept Maps, Flashcards and practice questions. Also free: the Domain 1 flashcards. New to the exam? Read the AIF-C01 study guide.