AWS AI Practitioner Exam

Prep Guide
Prepare for success with this comprehensive guide covering all
key domains of the AWS AI Practitioner exam. Master the
fundamentals of AI/ML, generative AI, foundation models,
responsible AI practices, and security requirements.
Domain 1: Fundamentals of AI and ML
Artificial Intelligence (AI) is the science of making machines intelligent, enabling them to perform tasks that
typically require human intelligence. Machine Learning (ML) is a subset of AI where algorithms learn patterns
from data without explicit programming.

Artificial Intelligence Machine Learning Deep Learning

The broad field enabling A subset of AI focused on A subset of ML using neural
machines to perform tasks algorithms that learn patterns networks with multiple layers
requiring human intelligence: from data without explicit to learn complex patterns from
reasoning, learning, and programming. large datasets.
problem-solving.
• Recommendation systems • Spam filters • Image recognition
• Voice assistants • Fraud detection • Natural language processing
• Self-driving cars • Predictive maintenance • Speech recognition
Key AI/ML Concepts
Neural Networks

Models inspired by the human brain, consisting of interconnected nodes

(neurons) that process information. Used in deep learning for complex pattern
recognition tasks.
Computer Vision

A field enabling computers to interpret and understand visual information from

the world, such as images and videos.

Natural Language Processing

Focuses on the interaction between computers and human language, enabling

machines to understand, interpret, and generate human language.
Training and Inference

Training Model Inference

The process of feeding data into a A mathematical representation of a The process of using a trained model to
model to adjust its parameters and real-world phenomenon, learned from make predictions or decisions on new,
learn patterns. Requires significant data. The trained model contains the unseen data. Often needs to be
computational resources and large patterns and relationships discovered optimized for performance and cost.
datasets. during training.

Types of Inference

Batch Inference Real-time Inference

• Processes data in batches • Processes data as it arrives

• Efficient for large datasets • Minimal latency
• Higher latency • Suitable for interactive applications
• Lower cost per prediction • Higher computational cost per prediction
Types of Machine Learning
Supervised Learning
The model learns from labeled data, where the correct output is
provided for each input.

Examples: Regression (predicting continuous values), Classification

(predicting categories)

Unsupervised Learning
The model learns patterns from unlabeled data without explicit guidance.

Examples: Clustering (grouping similar data), Dimensionality reduction

Reinforcement Learning
The model learns by interacting with an environment, receiving
rewards or penalties based on its actions.

Examples: Game playing, robotics, autonomous systems

Types of Data in AI Models

Tabular Data Time-Series Data

Data organized in rows and columns, similar to a spreadsheet. Each row Data collected over time, with a specific order and time stamps. Used to
represents a data point or record, and each column represents a specific analyze trends, make predictions, and detect anomalies.
feature or attribute.

Image Data Text Data

Visual data, such as photographs or digital images. Used in tasks like image Textual information, such as articles, documents, or social media posts. Used
classification, object detection, and image generation. in NLP tasks like sentiment analysis, text classification, and machine
translation.

Labeled vs. Unlabeled Data Structured vs. Unstructured Data

Labeled Data: Data points that have been manually assigned specific labels or Structured Data: Organized in a predefined format, such as databases or
categories. Essential for supervised learning. spreadsheets. Easy to analyze and process.

Unlabeled Data: Data points without associated labels. Used in unsupervised Unstructured Data: Data that lacks a specific format, such as text documents,
learning for pattern discovery. images, or videos. More challenging to process.
Practical AI/ML Use Cases
Identifying when AI/ML solutions are appropriate:

Assisting Human Decision Making Solution Scalability Automation

• Recommendation systems for • Natural language processing for • Robotic process automation for
personalized product suggestions multilingual customer support repetitive tasks
• Medical diagnosis tools analyzing • Computer vision for analyzing • Self-driving vehicles for transportation
medical images thousands of images • Intelligent document processing
• Financial analysis for predicting • Automated data processing for large systems
market trends datasets

When AI/ML May Not Be Appropriate

• When the cost of development outweighs potential benefits

• For simple deterministic tasks with clear rules
• When insufficient data is available for training
• When explainability and transparency are critical requirements
AWS Managed AI/ML Services

Amazon SageMaker
Fully managed service covering the entire ML lifecycle from data preparation to deployment. Provides built-in
algorithms, notebooks, and infrastructure for training and hosting.

Amazon Comprehend
Natural language processing service that uses machine learning to find insights and relationships in text.
Identifies entities, key phrases, sentiment, and more.

Amazon Rekognition
Image and video analysis service that can identify objects, people, text, scenes, and activities, as well as
detect inappropriate content.

Amazon Translate
Neural machine translation service that delivers fast, high-quality, and affordable language translation.
Supports dozens of languages.

Additional services include Amazon Transcribe (speech-to-text), Amazon Polly (text-to-speech), Amazon Lex
(conversational interfaces), and Amazon Textract (document analysis).
ML Development Lifecycle
Data Collection
Identify and gather relevant data from various sources. Ensure data is clean, accurate, and consistent by removing duplicates and handling missing values.

Exploratory Data Analysis

Gain insights into the data by visualizing distributions, correlations, and patterns. Discover relevant features and address outliers and inconsistencies.

Data Preprocessing
Handle missing values, create new features, normalize numerical features, and encode categorical features. Prepare data for model training.

Model Training & Tuning

Select appropriate algorithms, train models on prepared data, and optimize hyperparameters to improve model performance.

Model Evaluation
Assess model performance using relevant metrics and validate its ability to generalize to unseen data. Select the best-performing model.

Model Deployment
Deploy the model to a production environment and integrate it with other systems to make predictions.

Monitoring & Maintenance

Track performance over time, detect data drift, and retrain as needed to maintain accuracy and relevance.
Understanding MLOps
Machine Learning Operations (MLOps) bridges the gap between data science and engineering

Key MLOps Concepts

Experimentation: A/B testing, hyperparameter tuning, feature engineering

Repeatable Processes: Version control, pipeline orchestration, configuration
management
Scalable Systems: Cloud infrastructure, distributed training, model serving
Technical Debt Management: Clean code, modular design, refactoring
Production Readiness: Model deployment, monitoring, alerting
Model Monitoring: Performance tracking, data drift detection, concept drift monitoring
Model Re-training: Automated retraining pipelines, model versioning
Domain 2: Fundamentals of Generative AI
Generative AI creates new content like text, images, music, and code

Tokens, Chunking, and Embeddings Transformer Architecture

Tokens: Smallest units of text that AI models process Neural network architecture that excels at processing
(words, parts of words, characters) sequential data using self-attention mechanisms

Chunking: Breaking text into smaller sequences of Forms the foundation of Large Language Models (LLMs)
tokens for efficient processing like GPT and BERT

Embeddings: Numerical representations that capture

semantic meaning in vector space

Foundation Models Diffusion Models

AI models trained on massive datasets that can be Generative models that gradually add noise to an image
adapted to various tasks with minimal additional training and then learn to reverse this process

Examples include BERT, GPT, DALL-E, and Stable Diffusion Used for high-quality image generation, editing, and style
transfer
Generative AI Use Cases

Image & Video Generation Text Generation

• Creating realistic images from text descriptions • Content creation (articles, marketing copy)
• Style transfer between images • Text summarization and translation
• Generating video content and animations • Code generation and completion

Conversational AI Audio Generation

• Advanced chatbots and virtual assistants • Text-to-speech synthesis
• Customer service automation • Music composition
• Interactive storytelling experiences • Sound effects creation

Foundation Model Lifecycle

Data Selection 1
Gathering high-quality, diverse, and representative datasets

2 Pre-training
Training on massive datasets to learn general patterns and representations

Fine-tuning 3
Adapting the pre-trained model to specific tasks using smaller datasets

4 Evaluation
Assessing model performance against benchmarks and human evaluation

Deployment 5
Making the model available for use through APIs or applications

6
Advantages & Limitations of Generative AI
Advantages Limitations

Adaptability: Can be customized to specific needs and applied Hallucinations: Can generate incorrect or fabricated information
to various tasks Interpretability: Often operates as a "black box" with limited
Responsiveness: Produces content quickly, suitable for real- transparency
time applications Inaccuracy: May produce errors or irrelevant content,
Simplicity: Often offers intuitive interfaces, accessible to non- especially with ambiguous prompts
technical users Nondeterminism: Can produce different outputs for the same
Creativity: Generates novel and innovative content beyond input
human imagination Bias: May perpetuate or amplify biases present in training data
Efficiency: Automates content creation tasks that would take
humans much longer

Model Selection Factors

When choosing a generative AI model, consider: model type (GAN, VAE, LLM), performance requirements, capabilities
(customization, contextual understanding), resource constraints, and compliance requirements.
AWS Services for Generative AI
Amazon Bedrock SageMaker JumpStart
Fully managed service providing access to foundation Provides pre-trained models, solution templates, and
models from leading AI providers through a unified API. examples to help quickly get started with machine learning.
Includes models from Anthropic, AI21 Labs, Cohere, Meta, Includes foundation models that can be fine-tuned.
Stability AI, and Amazon.

Amazon Q PartyRock
Generative AI-powered assistant that can help with various Low-code/no-code playground for experimenting with
tasks like answering questions, generating content, and generative AI models and quickly creating AI-powered
providing code suggestions. applications without deep technical expertise.

Benefits of AWS Infrastructure for Generative AI

Security: Advanced security measures including encryption, Compliance: Services designed to comply with industry
access controls, and threat detection to protect models and data standards and regulations like GDPR, HIPAA, and PCI DSS

Responsibility: Tools and practices for responsible AI, including Scalability: Elastic infrastructure that scales to handle varying
bias mitigation and model safety features workloads efficiently
Domain 3: Applications of Foundation Models
Design Considerations for Foundation Model Applications

Model Selection Criteria

Cost: Licensing fees and inference costs

Modality: Text, image, audio, or multimodal capabilities

Latency: Response time requirements

Multi-lingual support: Languages needed for your application

Model size: Resource requirements vs. performance

Input/output length: Maximum context window

Retrieval Augmented Generation (RAG)
What is RAG?

RAG combines retrieval of relevant information from external knowledge sources with the generative
capabilities of foundation models. This approach:

• Enhances factual accuracy by grounding responses in verified information

• Reduces hallucinations by providing specific context
• Enables models to access current or domain-specific information
• Allows organizations to leverage their proprietary data

AWS Services for RAG Implementation

 Prompt Engineering Techniques
The art of crafting effective prompts to guide AI model outputs

Zero-Shot Prompting 1
Asking the model to perform a task without any examples

ample: "Translate the following sentence to Spanish: 'The weather is nice today.'"
2 Few-Shot Prompting
Providing a few examples to guide the model's understanding

Example: "Classify the sentiment: 'I love this product!' → Positive. 'This
Chain-of-Thought 3
didn't work at all.' → Negative. 'The service was okay.' →"
Breaking down complex problems into step-by-step reasoning

Example: "Let's solve this math problem step by step: If x+y=10 and x-
4 Role Prompting
y=4, what are the values of x and y?"
Assigning a specific role to the model to elicit particular expertise

Example: "As an experienced data scientist, explain how to handle

missing values in a dataset."

Prompt Engineering Risks

Prompt injection: Malicious users may attempt to override instructions

Data leakage: Sensitive information might be included in prompts
Jailbreaking: Attempts to bypass safety measures or guidelines

Effective prompt engineering requires clear instructions, specific context, and iterative refinement based on model responses. Always review and test prompts thoroughly.
Domain 4: Guidelines for Responsible AI
Developing AI systems that are ethical, fair, and beneficial

Key Features of Responsible AI AWS Tools for Responsible AI

Bias: Identifying and mitigating unfair prejudice in models Amazon SageMaker Clarify: Detects bias in models and
Fairness: Ensuring equitable treatment across different groups explains predictions

Inclusivity: Developing systems that work well for diverse Amazon Bedrock Guardrails: Sets boundaries for model behavior
populations SageMaker Model Cards: Documents model details,
Robustness: Building models resilient to unexpected inputs limitations, and ethical considerations

Safety: Protecting users from potential harms AWS Audit Manager: Assesses AI systems against
compliance frameworks
Veracity: Ensuring accuracy and truthfulness in outputs

Legal Risks of Generative AI

Be aware of potential risks including intellectual property infringement, perpetuating harmful biases, generating misleading
information, and privacy violations. Implement appropriate safeguards and governance structures.
Model Transparency and Explainability
Transparent vs. "Black Box" Models

Transparent models can provide clear explanations for their decisions, while "black box" models (like complex
deep neural networks) may be difficult to interpret despite high accuracy.

Trade-offs to Consider

Interpretability vs. Performance: Simpler, more interpretable models may sacrifice some accuracy

Privacy vs. Transparency: Detailed explanations might reveal sensitive information

Regulatory Requirements: Some industries require explainable AI for compliance

Domain 5: Security, Compliance, and Governance

Securing AI Systems Compliance Considerations Data Governance

IAM Roles and Policies: Control access to AI Regulatory Standards: ISO/IEC 27001, SOC Data Lifecycle Management: From
resources with least privilege principles 2, GDPR, HIPAA ingestion to retirement

Encryption: Protect data at rest and in Algorithm Accountability: EU AI Act, CCPA, Data Quality: Implement checks to identify
transit using AWS KMS and other emerging regulations issues

AWS PrivateLink: Establish private Data Residency: Store data in specific Data Retention: Define clear policies based
connections without exposure to the internet geographic regions as required on requirements

Data Logging: Track data flows and system Documentation: Source citation and data
events with CloudTrail provenance tracking

AWS Services for Governance and Compliance

AWS Config: Continuously monitors resource configurations Amazon Macie: Discovers and protects sensitive data
AWS Audit Manager: Simplifies compliance assessment AWS CloudTrail: Records API calls for auditing
AWS Artifact: Provides on-demand access to compliance reports AWS Trusted Advisor: Provides security recommendations

Remember that AWS follows a shared responsibility model: AWS secures the infrastructure, while you're responsible for securing your data, applications,
and access management.