AWS Certified AI Practitioner

Cheat Sheet
Quick Bytes for you before the exam!
The information provided in the Cheat Sheet is for educational purposes only; created in our efforts to help
aspirants prepare for the AWS AI Practitioner—Beta Exam. Though references have been taken from AWS
documentation, it’s not intended to be a substitute for the official docs. The document can be reused, reproduced,
and printed in any form; ensure that appropriate sources are credited and required permissions are received.

Are you Ready for

“AWS AI Practitioner” Certification?

Self-assess yourself with

Whizlabs FREE TEST

800+ Hands-on-Labs and Cloud Sandbox

Hands-on Labs Cloud Sandbox environments
`
 Index
Topics Names Page No
Fundamentals of AI and ML
Artificial Intelligence 4
Basic AI Terminologies 7
Differences between AI, ML, Deep Learning & Gen AI 9
Understanding Foundation Model 10
AI Models: Types 12
Machine Learning 14
ML Pipeline: Components with AWS Services 15
Fundamentals of ML Operations (MLOps) 17
Amazon SageMaker 19

Fundamentals of Generative AI
Generative AI 22
GenAI Security Scoping Matrix 25
Amazon SageMaker JumpStart 27
Amazon Bedrock 28
Amazon Q 31

Applications of Foundation Models

Prompt Engineering 32
Retrieval Augmented Generation (RAG) 34
RLHF - Reinforcement Learning from Human Feedback 36

Design and implement pipeline automation

Responsible AI 38
Amazon SageMaker Clarify 39
Amazon SageMaker Model Monitor 40

AWS Managed AI Services

Amazon Polly 42
Amazon Comprehend 42

—Back to Index— 2
Amazon Rekognition 43
Amazon Lex 44
Amazon Transcribe 45
Amazon Translate 46
Amazon Mechanical Turk (MTurk) 47
Amazon Augmented AI [Amazon A2I] 48
AWS DeepRacer 49

Security, Compliance, and Governance for AI Solutions

Amazon Macie 50
AWS PrivateLink 50

—Back to Index— 3
 Fundamentals of AI and ML
Artificial Intelligence
What is Artificial Intelligence?
● Artificial Intelligence (AI) is a field of computer science focused on developing systems
that can exhibit intelligent behavior, such as reasoning, learning, and autonomous
action.
● AI functions by combining large amounts of data with intelligent algorithms. These
algorithms are trained on the data to learn patterns and make informed decisions. This
process enables AI to perform tasks that require human-like intelligence, such as
understanding language and driving vehicles.
● AWS offers a range of pre-built AI services, as well as customizable infrastructure
options, designed to streamline AI development and reduce costs.
How does AI process information and make decisions?
Data Collection: AI systems require vast amounts of data to learn from. This data can be
anything from images and text to numerical values.
Algorithm Selection: The appropriate algorithm is chosen based on the specific task the AI is
designed to perform. Common algorithms include:
● Machine Learning: It involves training algorithms on large datasets to identify patterns
and make predictions.
● Deep Learning: A subset of machine learning that uses artificial neural networks to learn
from complex patterns.
● Natural Language Processing (NLP): AI can understand and respond to human language.
● Training: The algorithm is trained on the collected data. This involves feeding the data
into the algorithm and allowing it to learn from it.
● Testing: The trained algorithm is tested on new data to evaluate its performance.
● Deployment: Once the algorithm has been tested and refined, it can be deployed in
real-world applications.
Key components of AI application architecture:
Artificial intelligence architecture consists of three core layers. All the layers run on IT
infrastructure that provides the necessary compute and memory resources for the AI to run.
Layer 1: Data layer
The data layer is the cornerstone of AI. It involves preparing and organizing data for use in
machine learning, natural language processing, and image recognition technologies.

—Back to Index— 4
Layer 2: Model layer
The model layer focuses on the AI system's decision-making capabilities. Organizations often
select pre-existing foundation models or large language models and customize them using
techniques that incorporate relevant data.
Layer 3: Application layer
The application layer is the user-facing interface of the AI system. It allows users to interact with
the AI, request specific tasks, generate information, or make data-driven decisions.
Applications of Artificial Intelligence:
Chatbots and smart assistants-
● Chatbots and smart assistants powered by AI are increasingly capable of conducting
conversations that resemble human interactions.
Intelligent document processing-
● Intelligent document processing (IDP) transforms business documents such as emails,
images, and PDFs into organized, structured information. IDP leverages AI technologies,
including natural language processing (NLP), deep learning, and computer vision, to
extract, categorize, and validate data.
Application performance monitoring-
● AI-driven APM tools leverage historical data to anticipate potential issues before they
arise. Additionally, these tools can address problems in real time by offering developers
practical solutions, ensuring smooth application operations, and resolving performance
bottlenecks.
Predictive maintenance-
● AI-driven predictive maintenance leverages vast amounts of data to detect potential
issues that could cause operational, system, or service interruptions. By anticipating
problems before they arise, predictive maintenance helps businesses minimize
downtime and avoid disruptions.
Medical research-
● AI technology in medical research streamlines automates repetitive tasks, and handles
large datasets. It can be employed to enhance the entire pharmaceutical discovery and
development process, transcribe medical records, and accelerate the time-to-market for
new products.
Business Analytics-
● Business analytics leverages AI to gather, process, and examine intricate datasets. With
AI-driven analytics, you can predict future trends, identify underlying causes within the
data, and streamline time-consuming tasks.

—Back to Index— 5
Limitations of Artificial Intelligence in AWS:
While AWS offers a robust suite of AI and machine learning services, there are inherent
limitations that users should be aware of:
1. Data Quality and Quantity:
● Bias: If AI models are trained on biased data, it can lead to unfair or inaccurate results.
● Noise: Incomplete or noisy data can hamper model accuracy and performance.
2. Computational Resources:
● Cost: Training and running AI models can become expensive, especially for large-scale
applications.
● Infrastructure: Access to high-performance computing infrastructure may be required
for complex models.
3. Ethical Considerations:
● Privacy: Handling sensitive data raises privacy concerns, especially when using AI for
tasks like facial recognition or natural language processing.
● Fairness: AI models can perpetuate existing biases or discrimination if not designed and
trained carefully.
4. Human Oversight:
● Dependency: AI systems-related tools still require human oversight to ensure they are
used ethically and effectively.
● Error Correction: AI models can make mistakes, and human intervention may be
necessary to correct errors or biases.

—Back to Index— 6
 Basic AI Terminologies
What is Machine Learning?
Machine Learning (ML) is a subset of artificial intelligence (AI) focused on developing algorithms
that improve automatically through experience and data. Simply put, machine learning allows
computers to learn from data and make decisions or predictions without explicit programming.
What is Deep Learning?
Deep learning, a subset of artificial intelligence, enables computers to learn from data like human
cognition. By analyzing complex patterns within images, text, audio, and other forms of data, deep
learning models can provide accurate insights and predictions. These models can automate tasks
that traditionally necessitate human intelligence, such as image description or transcription of audio
files into text.
What is the Large Language Model [LLM])?
Large language models (LLMs) are sophisticated deep learning models trained on extensive text
datasets. These models utilize a transformer architecture, a neural network framework composed of
an encoder and decoder. The encoder and decoder leverage self-attention mechanisms to extract
contextual meaning from text sequences, comprehending the relationships between words and
phrases.
What is Responsible AI?
➔ Broader Implications: AI systems have substantial effects on individuals, communities,
and the environment.
➔ Ethical Priorities: Organizations should emphasize fairness, transparency, and ethical
considerations in AI practices.
➔ Balancing Act: Businesses must balance ethical AI practices with the pursuit of
competitive advantage in a fast-evolving field.
What is Neural networks?
Artificial neural networks serve as the foundation for many artificial intelligence systems. Inspired by
the human brain's structure and function, these networks employ interconnected computational
units, often referred to as artificial neurons or nodes. Similar to biological neurons, these nodes
process information through mathematical calculations. By working together in a network, these
nodes can collectively solve complex problems and learn from data.
What is Natural Language Processing (NLP)?
Natural language processing (NLP) employs neural networks to extract meaning from textual data. It
leverages computational methods designed to comprehend human language, enabling machines to
process words, grammar, and sentence structure. This technology facilitates tasks such as document
summarization, chatbot interactions, and sentiment analysis.

—Back to Index— 7
What is Computer vision?
Computer vision, powered by deep learning, enables computers to interpret and understand visual
information from images and videos. This technology can be applied in various domains, such as
content moderation to identify inappropriate images, facial recognition, and image classification. It
is also crucial in autonomous vehicles, where it helps to perceive the surrounding environment and
make timely decisions.
What is Speech Recognition?
Speech recognition technology leverages deep learning algorithms to decipher human speech,
discern individual words, and comprehend meaning. Neural networks are employed to transcribe
spoken language into written text and to gauge the emotional tone of the speaker. Speech
recognition finds applications in various domains, such as virtual assistants and call center systems,
where it is used to interpret spoken commands and execute corresponding actions.
What is Generative AI?
Generative AI is a type of artificial intelligence that can produce original content, including images,
videos, text, and audio, based on textual instructions. Unlike traditional AI, which primarily analyzes
data, generative AI utilizes deep learning techniques and vast datasets to create high-quality,
human-like creative outputs.
While this technology opens up new possibilities for creative expression, there are also concerns
regarding bias, harmful content, and intellectual property rights. In conclusion, generative AI marks
a significant advancement in AI's ability to generate human-quality language and content.

—Back to Index— 8
 Differences between AI, ML, Deep Learning & Gen AI
Aspect Artificial Intelligence (AI) Machine Learning (ML) Deep Learning Generative AI

Definition AI is a broad field of ML is a subset of AI DL is a subset of Gen AI is a type of AI

creating machines focused on systems that ML that uses focused on
capable of performing learn from data to make neural networks generating new
tasks that typically decisions. with multiple content, such as
require human layers (deep neural images, text, or
intelligence. networks) to learn music, based on
from large learned patterns.
amounts of data.

Types Rule-based systems, Supervised, unsupervised, Convolutional Generative

expert systems, decision and reinforcement neural networks Adversarial Networks
trees, ML, and DL. learning. (CNNs), recurrent (GANs), Variational
neural networks Autoencoders (VAEs),
(RNNs), Large Language
transformers. Models (LLMs).

Data It can include rule-based It relies on labeled or It requires large It requires large
Dependency systems with no learning unlabeled data for learning amounts of data to datasets for training
component. patterns. train deep neural to generate realistic
networks outputs.
effectively.

Applications Robotics, natural Predictive analytics, Image recognition, Image and video
language processing, recommendation systems, NLP, autonomous generation, text
expert systems, etc. fraud detection. vehicles. generation (e.g.,
chatbots), music
composition.

Examples Siri, autonomous robots, Spam filters, personalized Facial recognition DALL-E for image
chess-playing computers. recommendations, systems, generation, ChatGPT
weather forecasting. self-driving cars, for text generation,
and language and DeepDream for
translation image creation.
services.

—Back to Index— 9
 Understanding Foundation Model
What Are Foundation Models?
Foundation models (FMs) are large-scale deep-learning neural networks trained on extensive
datasets. They have revolutionized the approach data scientists take to machine learning (ML)
by providing a starting point that speeds up the development of new AI applications. These
models are designed to perform a broad range of general tasks, such as language
understanding, text and image generation, and natural language processing (NLP).
Features of Foundation Models
Adaptability:
Foundation models are uniquely versatile and capable of executing a variety of tasks with high
accuracy based on input prompts. This makes them significantly different from traditional ML
models, which are typically designed for specific tasks like sentiment analysis, image
classification, or trend forecasting.
General-Purpose Nature:
Due to their large size and broad training, foundation models can serve as base models for more
specialized applications. Over the years, these models have grown in complexity and size, with
models like BERT and GPT-4 showcasing this evolution.
Applications of Foundation Models
Language Processing:
Foundation models excel in natural language tasks, including answering questions, writing
scripts, and translating languages.
Visual Comprehension:
These models are highly effective in computer vision, identifying images, generating images
from text, and editing photos and videos.
Code Generation:
Foundation models can write and debug code in various programming languages based on
natural language instructions.
Human-Centered Engagement:
They support decision-making processes, such as clinical diagnoses and analytics, by
continuously learning from human inputs during inference.
Examples of Foundation Models
BERT (2018):
A bidirectional model trained on a vast dataset, capable of analyzing text and predicting
sentences. It laid the groundwork for future models like GPT.
GPT (Generative Pre-trained Transformer):
Released by OpenAI, GPT models have evolved from GPT-1 with 117 million parameters to

—Back to Index— 10
GPT-4, which boasts 170 trillion parameters. These models are capable of tasks ranging from
text generation to question answering.
Amazon Titan:
A foundation model from Amazon offers generative and embedding models for tasks like text
summarization, information extraction, and personalization.
Challenges with Foundation Models
High Resource Demands:
Developing foundation models requires substantial infrastructure, making it costly and
time-intensive.
Integration Complexity:
For practical use, these models must be integrated into software systems, which involves
additional development for prompt engineering and fine-tuning.
Comprehension and Reliability Issues:
While foundation models can generate coherent responses, they may struggle with
understanding context and can produce unreliable or biased answers.
AWS Support for Foundation Models
Amazon Bedrock:
This service simplifies the development and scaling of generative AI applications by offering
access to foundation models via an API, allowing users to choose the most suitable model for
their needs.
Amazon SageMaker JumpStart:
A hub for ML models and solutions, SageMaker JumpStart provides access to a wide range of
foundation models, including popular ones like Llama 2 and Falcon, supporting the
development of diverse AI applications.
Reference:
https://aws.amazon.com/what-is/foundation-models/

—Back to Index— 11
 AI Models: Types
1. Computer Vision Models
● Amazon Rekognition: This service provides pre-trained models for image and video
analysis, including capabilities like object detection, facial recognition, and scene
detection.
2. Natural Language Processing (NLP) Models
● Amazon Comprehend: Used for analyzing text, Amazon Comprehend can perform
sentiment analysis, entity recognition, and language detection.
● Amazon Translate: Provides real-time translation services between different languages.
● Amazon Lex: Powers conversational interfaces, enabling the creation of chatbots that
can interact through voice and text.
● Amazon Polly: Converts written text into lifelike speech in multiple languages.
3. Speech Recognition Models
● Amazon Transcribe: This service converts speech into text, making it useful for
transcriptions, subtitles, and more.
4. Document Processing Models
● Amazon Textract: Extracts text, tables, and other data from scanned documents, making
it easier to process and analyze paper-based information.
5. Recommendation and Forecasting Models
● Amazon Personalize: Delivers personalized recommendations by analyzing user behavior
and preferences.
● Amazon Forecast: Utilizes time-series data to predict future trends, such as sales
forecasts or inventory needs.
6. Search and Information Retrieval Models
● Amazon Kendra: An enterprise search service that uses machine learning to provide
relevant search results across documents and data sources.
7. Custom Machine Learning Models
● Amazon SageMaker: A comprehensive platform for building, training, and deploying
custom machine learning models. It supports a wide range of algorithms and
frameworks.
8. Generative AI Models
● Amazon Bedrock: A service that provides access to foundational models for generative
AI, allowing users to create custom applications like text generation or image creation.
● SageMaker JumpStart: Offers pre-trained models and solutions for generative AI tasks,
which can be fine-tuned for specific needs.

—Back to Index— 12
9. Edge AI Models
● AWS IoT Greengrass ML Inference: Enables machine learning inference on edge devices,
allowing models to be deployed in environments where real-time processing is critical.
10. Hybrid AI Models
● Amazon Neptune ML: Integrates machine learning with graph databases, enabling
advanced data analysis and knowledge graph applications.

—Back to Index— 13
 Machine Learning
What is Machine Learning?
● Core Concept: Machine learning revolves around creating algorithms that facilitate
decision-making and predictions. These algorithms enhance their performance over
time by processing more data.
● Traditional vs. ML Programming: Unlike traditional programming, where a computer
follows predefined instructions, machine learning involves providing a set of examples
(data) and a task. The computer then figures out how to accomplish the task based on
these examples.
● Example: To teach a computer to recognize images of cats, we don’t give it specific
instructions. Instead, we provide thousands of cat images and let the machine learning
algorithm identify common patterns and features. Over time, the algorithm improves
and can recognize cats in new images it hasn’t seen before.
Types of Machine Learning
Machine learning can be broadly classified into three types:
1. Supervised Learning: The algorithm is trained on labeled data, allowing it to make
predictions based on input-output pairs.
2. Unsupervised Learning: The algorithm discovers patterns and relationships within
unlabeled data.
3. Reinforcement Learning: The algorithm learns by trial and error, receiving feedback
based on its actions.
Applications of Machine Learning
Machine learning powers many of today’s technological advancements:
● Voice Assistants: Personal assistants like Siri and Alexa rely on ML to understand and
respond to user queries.
● Recommendation Systems: Platforms like Netflix and Amazon use ML to suggest content
and products based on user behaviour.
● Self-Driving Cars: Autonomous vehicles use ML to navigate and make real-time
decisions.
● Predictive Analytics: Businesses use ML to forecast trends and make data-driven
decisions.

—Back to Index— 14
 ML Pipeline: Components with AWS Services
A Machine Learning (ML) pipeline in AWS refers to a structured workflow that automates the
various stages involved in developing, training, and deploying machine learning models.
1. Data Collection
● Amazon S3 (Simple Storage Service): Used to store large datasets. AWS provides secure
and scalable storage for structured and unstructured data.
● AWS Glue: A data integration service that helps to discover, prepare, and combine data
across multiple sources for analysis.
● Amazon RDS (Relational Database Service): For storing and managing relational data
that can be used for training ML models.
2. Exploratory Data Analysis (EDA)
● Amazon SageMaker Studio: Provides an integrated environment where data scientists
can perform EDA using Jupyter notebooks. It supports visualization libraries like
Matplotlib, Seaborn, and Pandas for statistical analysis and data exploration.
● Amazon Athena: An interactive query service that allows you to analyze data in Amazon
S3 using SQL. Useful for quick analysis without the need to move data.
3. Data Pre-processing
● AWS Glue and AWS Data Wrangler: These tools help in cleaning, normalizing, and
transforming raw data into a format suitable for modeling. This may involve handling
missing values, normalization, and data scaling.
● Amazon SageMaker Processing: Allows running pre-processing jobs that can scale to
handle large datasets.
4. Feature Engineering
● Amazon SageMaker Feature Store: A fully managed repository for storing, retrieving,
and sharing features across different models and teams. It helps in automating the
process of feature extraction and management.
● Amazon SageMaker Data Wrangler: Simplifies the process of feature transformation,
enabling users to create new features by combining existing ones.
5. Model Training
● Amazon SageMaker: Supports training custom models using a wide variety of built-in
algorithms or your own code. It also offers distributed training, enabling you to scale
training jobs across multiple instances.
● AWS Deep Learning AMIs: Provides pre-configured environments with popular deep
learning frameworks like TensorFlow, PyTorch, and Apache MXNet.

—Back to Index— 15
6. Hyperparameter Tuning
● Amazon SageMaker Automatic Model Tuning: Also known as hyperparameter
optimization (HPO), this service automatically tunes the model’s hyperparameters to
improve performance, using techniques like Bayesian optimization.
7. Model Evaluation
● Amazon SageMaker Debugger: Offers insights into the training process by monitoring
and profiling training jobs. It helps in identifying issues like overfitting and underfitting
by analyzing training metrics.
● Amazon SageMaker Model Monitor: Used post-deployment to track model
performance and detect data drift over time, ensuring that the model remains accurate.
8. Model Deployment
● Amazon SageMaker Endpoint: Allows you to deploy your trained models in real-time,
making them accessible via API for inference.
● Amazon Elastic Kubernetes Service (EKS): Supports deploying models in a
Kubernetes-managed environment for larger, more complex applications.
9. Monitoring
● Amazon CloudWatch: Monitors deployed models in real-time, collecting and tracking
metrics, logging, and triggering alerts for model performance or infrastructure issues.
● Amazon SageMaker Model Monitor: Continuously monitors deployed models for
concept drift, data quality issues, and other anomalies that might affect the model’s
accuracy over time.

—Back to Index— 16
 Fundamentals of ML Operations (MLOps)
MLOps in AWS is a set of practices that combine Machine Learning (ML) and DevOps to
streamline the development, deployment, and management of ML models in the Amazon Web
Services (AWS) cloud environment.
1. Experimentation
● Rapid Prototyping: AWS services like Amazon SageMaker allow data scientists to quickly
build, test, and iterate on machine learning models using Jupyter notebooks and
pre-built algorithms.
● Experiment Tracking: SageMaker Experiments helps in tracking and comparing different
model runs, capturing parameters, configurations, and outcomes for better
reproducibility and collaboration.
2. Repeatable Processes
● Pipeline Automation: SageMaker Pipelines automates the entire machine learning
workflow, from data preparation to model deployment, ensuring that each step is
repeatable and consistent.
● Infrastructure as Code (IaC): Using AWS CloudFormation or Terraform, you can define
and deploy infrastructure in a consistent and repeatable manner, ensuring that
environments are identical across different stages.
3. Scalable Systems
● Elastic Resources: AWS provides scalable compute resources like EC2 instances and
SageMaker-managed instances that can automatically scale up or down based on the
workload, ensuring efficient use of resources.
● Distributed Training: SageMaker supports distributed training, allowing large-scale
models to be trained faster across multiple GPUs or instances.
4. Managing Technical Debt
● Version Control: Versioning models, datasets, and code ensures that you can track
changes, reproduce results, and avoid issues caused by outdated or inconsistent
components.
● Model Registry: SageMaker Model Registry helps in managing different versions of
models, storing metadata, and promoting models through various stages of
development and production.
5. Achieving Production Readiness
● Continuous Integration/Continuous Deployment (CI/CD): Implementing CI/CD pipelines
with AWS CodePipeline or Jenkins integrates code changes, tests, and deployments
seamlessly, ensuring models are always production-ready.

—Back to Index— 17
 ● Security and Compliance: AWS provides tools like AWS Identity and Access Management
(IAM) and AWS Key Management Service (KMS) to secure data, models, and pipelines,
ensuring compliance with industry standards.
6. Model Monitoring
● Performance Monitoring: SageMaker Model Monitor automatically monitors deployed
models for accuracy and performance drift, alerting teams to any issues that may
require attention.
● Logging and Analytics: AWS CloudWatch and AWS X-Ray can be used to log model
predictions, track performance metrics, and diagnose issues in real time.
7. Model Re-training
● Automated Retraining: SageMaker Pipelines and Step Functions can automate the
retraining process when a model's performance drops or new data becomes available.
● Data Drift Detection: Monitoring tools like SageMaker Model Monitor can detect when
input data distribution shifts, triggering a model retraining pipeline to ensure the model
remains accurate.
8. Scalability and Flexibility
● Scalable Deployment: SageMaker endpoints can automatically scale to handle
increasing traffic, ensuring that the model can serve predictions efficiently regardless of
load.
● Multi-Model Endpoints: Allows deploying multiple models on a single endpoint,
optimizing resource utilization and reducing costs.
9. Collaboration and Governance
● Collaboration Tools: SageMaker Studio provides a unified interface where data scientists
and engineers can collaborate, share experiments, and work on models together.
● Governance and Auditing: AWS provides tools to maintain governance, such as
SageMaker Clarify for bias detection and SageMaker Model Monitor for ensuring model
compliance with business rules.
10. Technical Debt Management
● Artifact Management: Using services like S3 for storing datasets, models, and logs helps
in managing and organizing artifacts efficiently, reducing the technical debt associated
with disorganized resources.
● Code Reusability: Utilizing modular code and standardized practices across teams
minimizes redundant work and accelerates future projects.

—Back to Index— 18
 Amazon SageMaker
What is Amazon SageMaker?
● Amazon SageMaker is a comprehensive platform that empowers users to develop, train,
and deploy machine learning models efficiently. This fully managed service offers a wide
range of tools, including notebooks, debuggers, profilers, pipelines, and MLOps
capabilities, to streamline the entire ML lifecycle.
● Amazon SageMaker offers a range of pre-built tools, including algorithms, pre-trained
models, and solution templates, to expedite the development and deployment of
machine learning models to help data scientists and practitioners.

Algorithm Selection:
Problem Type Appropriate Algorithm

Binary Classification Logistic Regression, XGBoost, etc.

Multiclass Classification XGBoost, Linear Learner, etc.

Regression Linear Learner, XGBoost, etc.

Object Detection Faster R-CNN, SSD, etc.

Anomaly Detection Random Cut Forest, etc.

Clustering K-Means, DBSCAN, etc.

Topic Modeling Latent Dirichlet Allocation (LDA)

Recommender Systems Factorization Machines, etc.

Features:
Prepare Data -
● SageMaker Feature Store:- Amazon SageMaker Feature Store is a centralized platform
designed to store, share, and manage features used in machine learning models.
Features are the data inputs that models rely on during both training and inference.
● SageMaker Data Wrangler:- Amazon SageMaker Data Wrangler selects, understands,
and transforms data to prepare it for machine learning (ML) in minutes. reduces data
prep time for tabular, image, and text data from weeks to minutes. It enables a rapid
assessment of ML model accuracy and helps identify potential problems before
deployment.

—Back to Index— 19
 ● Geospatial ML with Amazon SageMaker:- Amazon SageMaker empowers data scientists
and ML engineers to build, train, and deploy ML models using geospatial data such as
satellite imagery, maps, and location data.
Build -
● SageMaker Notebooks:- Amazon SageMaker Notebooks offer a fully managed Jupyter
environment, enabling data scientists and ML engineers to explore, analyze, and develop
machine learning models efficiently.
● SageMaker Jumpstart:- Amazon SageMaker JumpStart is a machine learning (ML) hub
that can help you quickly evaluate, compare, and select Foundation models based on
pre-defined quality and responsibility metrics to perform tasks like article summarization
and image generation.
● SageMaker Studio Lab:- Amazon SageMaker Studio Lab is a free service based on
open-source JupyterLab that allows customers to use AWS compute resources to create
and run their Jupyter notebooks.
Train -
● SageMaker Model Training:- Amazon SageMaker Model Training streamlines the
process of training and tuning machine learning models, significantly reducing time and
costs while eliminating the need for infrastructure management.
● SageMaker Experiments:- Amazon SageMaker offers a managed MLflow capability that
simplifies machine learning and generative AI experimentation. Data scientists can easily
use MLflow within SageMaker for model training, registration, and deployment.
Administrators can quickly establish secure and scalable MLflow environments on AWS.
● SageMaker HyperPod:- Amazon SageMaker HyperPod simplifies the process of building
and optimizing ML infrastructure for training foundation models, significantly reducing
training time by 40%. By automatically distributing training workloads across thousands
of accelerators, HyperPod enables parallel processing and accelerates model
performance.
Deploy -
SageMaker Model Deployment:- Amazon SageMaker simplifies the deployment of machine
learning models, including foundation models, offering optimal cost-effectiveness for inference
requests across various applications.
● SageMaker Pipelines:- Amazon SageMaker Pipelines is a serverless workflow
orchestration service that automates machine learning (ML) and large language model
(LLM) workflows.

—Back to Index— 20
End-to-End ML -
● SageMaker MLOps:- Amazon SageMaker offers specialized tools for managing machine
learning operations (MLOps), streamlining and standardizing processes throughout the
machine learning lifecycle.
● SageMaker Canvas:- Amazon SageMaker Canvas offers a visual interface that simplifies
the machine learning process. It allows you to prepare data, build, and deploy ML
models efficiently.
● SageMaker Studio:- Amazon SageMaker Studio provides a comprehensive suite of tools
for the entire machine learning development steps, including data preparation, model
building, training, deployment, and management.

Feature SageMaker Canvas SageMaker Studio

Target Data scientists and ML engineers with Data scientists and ML engineers with
Audience limited coding experience advanced coding skills
Integrated development environment
Interface Visual, no-code interface (IDE)
Model Automated model selection and Manual model selection and training
Building training using various algorithms
Basic MLOps features (monitoring, Advanced MLOps capabilities (pipeline
MLOps versioning) creation, experiment tracking)
Rapid prototyping, exploratory data Complex ML models, custom pipelines,
Use Cases analysis, simple ML models research projects

SageMaker Ground Truth:- Amazon SageMaker Ground Truth provides a robust platform for
incorporating human expertise into the machine learning process, enhancing model
performance through continuous feedback.
ML Governance -
● ML Governance with SageMaker:- Amazon SageMaker offers specialized governance
features to ensure responsible machine-learning practices. Amazon SageMaker Role
Manager allows administrators to quickly establish necessary permissions.
● SageMaker Clarify:- SageMaker Clarify streamlines the process of identifying potential
biases in your dataset. Specify the input features you're concerned about, like gender or
age, and SageMaker Clarify will conduct a thorough analysis to uncover any potential
biases present in those features.

—Back to Index— 21
 Fundamentals of Generative AI
Generative AI
What is Generative AI?
● Generative AI is a form of artificial intelligence capable of generating original content,
such as text, visuals, and audio.
● You can train it to understand and generate text in human and programming languages.
It can also learn complex subjects like art, chemistry, biology, etc. By leveraging past
training data, it can apply its knowledge to new and unfamiliar tasks.
● Using AWS, you can rapidly develop and expand generative AI applications tailored to
your specific data, use cases, and customer needs. Benefit from enterprise-grade
security and privacy, access to cutting-edge foundational models, and a data-centric
approach.
How does Generative AI work?
Generative AI utilizes machine learning models, which are extensively trained on massive
datasets, to produce new content.
Foundation models:- Foundation models are machine learning models trained on massive,
diverse datasets without specific task labels. They are capable of performing a wide variety of
general tasks.
Large language models:- LLMs are one class of FMs. These models are capable of performing a
variety of linguistic functions, including summarization, text generation, classification,
open-ended dialogue, and information extraction.
Benefits of Generative AI:
● Accelerates research
● Enhances customer experience
● Optimizes business processes
● Boosts employee productivity
Generative AI Models:
Diffusion models:- Diffusion models create new data by gradually introducing noise to existing
data samples and then carefully removing it. This process involves adding controlled random
changes to the original data over multiple iterations. The model ensures that the generated
data remains coherent and realistic by carefully controlling the noise level.
Once the data has been sufficiently noised, the diffusion model reverses the process. It
gradually removes the noise, step by step, until it produces a new data sample that closely
resembles the original. This reverse denoising process allows the model to generate new data
that is consistent with the underlying data distribution.

—Back to Index— 22
Generative Adversarial Networks (GAN):- GANs are widely used in generating realistic images,
transforming visual styles, and data augmentation tasks.
Generative Adversarial Networks (GANs) employ a competitive training process between two
neural networks. The generator network produces fake data by introducing random noise, while
the discriminator network attempts to differentiate between this fake data and real data.
As training progresses, the generator refines its ability to create increasingly realistic data, while
the discriminator becomes more adept at identifying fake data. This iterative process continues
until the generator generates data that is indistinguishable from authentic data, even to the
discriminator.
Variational Auto-encoders:- VAEs create a compressed representation of data, often referred to
as a latent space. This latent space is a mathematical construct that captures the essence of the
data. Imagine it as a unique code, summarizing the key features of the data.
For example, if studying faces, the latent space contains numbers representing eye shape, nose
shape, cheekbones, and ears.
VAEs use two neural networks. The encoder network transforms input data into a mean and
variance for each dimension of the latent space. A random sample is drawn from a Gaussian
distribution using these parameters, resulting in a latent space representation. This compressed
representation serves as a simplified version of the input. The decoder network takes this latent
sample and attempts to reconstruct the original data. The quality of the reconstruction is
evaluated using mathematical metrics that compare the reconstructed output to the original
input.
Transformer-based models:- The transformer-based generative AI model builds upon the
encoder and decoder concepts of VAEs.
Transformer-based models use a self-attention mechanism. Self-attention helps enable
generative models to prioritize significant words during processing. Transformer-based models
leverage multiple multiple encoder layers called attention heads to identify diverse
interconnections between words. Each head focuses on distinct sections of the input sequence,
facilitating a comprehensive analysis of the data.
Tools for building generative AI applications:
● Amazon Bedrock
● Amazon SageMaker
● AWS Trainium
● AWS Inferentia
● Amazon EC2 P5 Instances
● Amazon EC2 UltraClusters

—Back to Index— 23
Generative AI-powered applications:
● Amazon Q
● PartyRock - Amazon Bedrock Playground
● AWS HealthScribe
Use Cases of Generative AI:
● Chatbots and Virtual Assistants
● Conversational Analytics
● Employee Assistant
● Code Generation
● Personalization
● Productivity and Creativity solution templates
Limitations of Generative AI on AWS:
Security:- The use of proprietary data to train generative AI models raises concerns about data
privacy and security. It's crucial to implement measures that protect sensitive information and
prevent unauthorized access.
Creativity:- While generative AI can produce creative content, its outputs are often constrained
by the data it has been trained on. Human creativity, which involves deeper understanding,
emotional resonance, and original thought, remains a challenge for AI to fully replicate.
Cost:- Training and running generative AI models require significant computational resources.
Cloud-based solutions offer a more accessible and cost-effective approach compared to building
models from scratch.
Explainability:- Generative AI models are often considered "black boxes," making it difficult to
understand their decision-making processes. Improving their interpretability and transparency
is essential to build trust and promote wider adoption.

—Back to Index— 24
 GenAI Security Scoping Matrix
A Generative AI Security Scoping Matrix offers a structured approach for organizations to
evaluate and implement security measures across the entire lifecycle of AI applications. By
categorizing security concerns, it provides a targeted framework for safeguarding AI systems.
● AWS provides multiple services to secure Generative AI workloads. AWS services vary
significantly in their underlying infrastructure, software, access mechanisms, and data
handling. To simplify security management, we've organized these services into logical
categories called 'scopes'.
Scoping (Determine your scope):
● To begin, you'll need to determine which scope your use case fits into.
● The scopes are numbered 1–5, representing the least ownership to greatest ownership
your organization has over the AI model and its associated data.

Figure 1: Generative AI Security Scoping Matrix

Buying generative AI:
● Scope 1: Consumer app – Your business consumes a public third-party generative AI
service, that is either free or paid. At this scope, you do not have ownership or access to
the underlying training data or model. You can only interact with the service through its
provided APIs or applications, adhering to the provider's terms of use.
Example: A worker uses a generative AI chatbot to brainstorm marketing campaign
concepts.
● Scope 2: Enterprise app – Your business uses a third-party enterprise application that
has generative AI capabilities, and a business relationship is established between your
organization and the vendor.

—Back to Index— 25
 Example: You use a third-party enterprise scheduling application that has a generative AI
capability embedded within to help draft meeting agendas.
● Scope 3: Pre-trained models – Your business utilizes an existing third-party generative AI
foundation model to power its application. This model is accessed and integrated into
your operations via an application programming interface.
Example: A customer support chatbot was developed utilizing the Anthropic Claude
foundation model, accessed via the Amazon Bedrock API.
● Scope 4: Fine-tuned models – Your business refines an existing third-party generative AI
foundation model by fine-tuning it with data specific to your business, generating a new,
enhanced model that’s specialized to your workload.
Example: By leveraging a foundation model through an API, you can create a marketing
application that tailors promotional materials specifically to your products and services.
● Scope 5: Self-trained models – Your business builds and trains a generative AI model
from scratch using data that you own or acquire. You own every aspect of the model.
Example: Your business wants to create a model trained exclusively on deep,
industry-specific data to license companies in that industry, creating a completely novel
LLM.
By identifying the specific applications of generative AI, security teams can prioritize their
efforts and evaluate the potential risks within each security domain.
Let's examine how scoping influences security requirements within each security
discipline.
● Governance and compliance – Implementing effective policies, procedures, and
reporting mechanisms can enable businesses to operate efficiently while mitigating
risks.
● Legal and privacy – The specific legal, regulatory, and privacy requirements for using or
creating generative AI solutions.
● Risk management – Assessing risks associated with generative AI and proposing
countermeasures.
● Controls – Implementing security measures to reduce risk.
● Resilience – Designing reliable generative AI systems that consistently meet business
SLAs.

—Back to Index— 26
 Amazon SageMaker JumpStart
Amazon SageMaker JumpStart is a machine learning hub that can speed up your ML
development. Using SageMaker JumpStart, you can select, evaluate, and compare FMs quickly
based on pre-defined quality and responsibility metrics to perform tasks like image generation
and article summarization.
Features:
Foundation Models:- Discover a variety of foundational models from leading providers like AI21
Labs, Databricks, Hugging Face, Meta, Mistral AI, Stability AI, and Alexa. These models can be
used to accomplish a wide range of tasks including summarizing articles and generating text,
images, or videos.
Built-in algorithms:- You can utilize built-in solution templates through the SageMaker Python
SDK. These algorithms address common ML tasks, including image, text, and tabular data
classification, as well as sentiment analysis.
Prebuilt solutions:- SageMaker JumpStart offers pre-built, end-to-end solutions for common
machine learning applications like demand forecasting, credit risk assessment, fraud detection,
and computer vision.
Benefits of SageMaker JumpStart:
Publicly available foundation models
Built-in ML algorithms
Customizable solutions
Support collaboration
Use cases and Advantages:
1. Foundation Model Integration
● Deploy models like LLaMA 2 and Stable Diffusion in VPC mode, even without internet.
● Access pre-trained models for easy deployment and tuning.
2. Large Language Models (LLMs)
● Simplifies deploying and tuning LLMs, including 40M parameter models for NLP tasks.
3. Text Classification
● Pre-built models for text classification with customization options.
4. Image Generation
● Deploy Stable Diffusion XL for high-quality image generation.
5. No-Code Solutions
● Fast, no-code deployment for quick AI solutions, accessible to non-experts.
6. Learning Resources
● Video tutorials and guides for easy model deployment and tuning.

—Back to Index— 27
 Amazon Bedrock
● Amazon Bedrock is a managed serverless service providing various high-performing
foundation models (FMs) from top AI companies like AI21 Labs, Anthropic, Cohere,
Meta, Mistral AI, Stability AI, and Amazon.
● These models are accessible via a unified API for creating generative AI applications
focusing on security, privacy, and responsible AI practices.
● With Amazon Bedrock, you can quickly test and compare different foundation models to
find the best fit for your use case.
● These models can be tailored to your unique data using techniques like fine-tuning and
Retrieval Augmented Generation (RAG).
● Additionally, you can create agents that can perform tasks using your company's systems
and information.
How does Amazon Bedrock help to build generative AI applications?
● Model Choice - Choose from a range of leading FMs: Amazon Bedrock’s single API lets
you easily switch between different foundation models and their updates.
● Customization - Privately adapt models with your data: Model customization lets you
deliver differentiated and personalized user experiences. Fine-tune foundation models
with your data to create unique, personalized experiences.
● RAG - Deliver more relevant FM responses: To provide FMs with relevant company data,
organizations use RAG. This technique feeds data into prompts to improve responses.
● Agents - Execute complex tasks across company systems: Amazon Bedrock agents
automate complex tasks using your company's systems and data. Agents analyze
requests, execute relevant APIs, and provide secure, private responses.
Amazon Bedrock offers models in 3 states:
● Active: The model provider is actively developing this version, and it will continue to be
updated with bug fixes and minor improvements.
● Legacy: A version is marked as a legacy when a more advanced version delivers superior
results. Amazon Bedrock determines an EOL date for outdated versions.
● EOL: This version is outdated and inoperable. Requests made to it will fail.
Use cases:
● Text generation - Produce unique content for your blog, social media, and web pages
● Virtual assistants - Build assistants that understand user inquiries, automatically divide
tasks, interact conversationally to gather necessary details, and execute actions to
complete the requested task.
● Text and image search - Identify and compile relevant information to answer questions
and provide recommendations based on a large body of textual and visual data.

—Back to Index— 28
 ● Text summarization - Acquire concise summaries of extensive documents, such as
articles, reports, research papers, technical documentation, and even books, to
effectively extract essential information.
● Image generation - Generate lifelike and visually engaging images for advertising
campaigns, websites, presentations, and other applications.
Amazon Bedrock Agents:
Amazon Bedrock Agents enable you to develop and configure autonomous agents for your
application.
Features:
● Amazon Bedrock Agents securely access your company's data, enhance user requests
with relevant information, and provide accurate responses.
● Amazon Bedrock Agents orchestrate and analyze the tasks, dividing them into the
appropriate logical order using the FM's reasoning capabilities.
● Amazon Bedrock Agents enable the dynamic generation and execution of code in a
secure environment. This automates complex analytical queries previously difficult to
address using model reasoning alone.
● Amazon Bedrock Agents enable you to incorporate business logic into your chosen
backend service. Moreover, the return of control functionality empowers you to execute
time-consuming actions in the background (asynchronously) while continuing the
orchestration flow.
● Amazon Bedrock Agents possess the ability to maintain memory across interactions,
allowing them to remember historical conversations and improve the accuracy of
multi-step tasks.
Amazon Bedrock Guardrails:
Amazon Bedrock Guardrails helps you establish safeguards for generative AI applications aligned
with your specific use cases and responsible AI policies.
Amazon Bedrock Guardrails offer additional customizable protections beyond the built-in
safeguards of foundation models.
Amazon Bedrock Guardrails protects your generative AI applications by evaluating both user
prompts and model responses.
● Blocking up to 85% more harmful content.
● Filtering out over 75% of inaccurate responses for RAG and summarization tasks.
● Customers can customize and implement safety, privacy, and truthfulness safeguards
within a unified solution.

—Back to Index— 29
Features:
● Amazon Bedrock Guardrails can be combined with Amazon Bedrock Agents and
Knowledge Bases to create generative AI applications that adhere to your responsible AI
policies.
● Customers can establish multiple guardrails, each tailored with a different combination
of controls, and apply these guardrails to various applications and use cases.
● Amazon Bedrock Guardrails offers customizable content filters to screen for harmful
content, including hate speech, insults, sexually suggestive material, violence,
misconduct (including criminal activity), and prompt attacks (prompt injection and
jailbreaking).
● Amazon Bedrock Guardrails employs contextual grounding checks to identify and filter
hallucinations when responses deviate from the provided information, such as being
factually incorrect or introducing new data.
PartyRock - Amazon Bedrock Playground:
PartyRock is a powerful tool designed to let you explore and experiment with the various
foundation models available on the Amazon Bedrock platform. It is designed specifically for
entertainment and creativity.
Features:
● Chat playground - The chat playground allows you to interact with the conversational
models available on Amazon Bedrock. When you enter a prompt into the model
● Text playground - The text playground offers a platform to explore Amazon Bedrock's
text models. By inputting a text prompt, you can see the model's generated output.
● Image playground - The image playground allows you to explore the capabilities of
Amazon Bedrock's image models. By entering a text description, you can see how the
model transforms your words into a visual representation.

—Back to Index— 30
 Amazon Q
Amazon Q is a generative AI-powered assistant that helps accelerate software development and
uses companies’ internal data.

Amazon Q Business is a generative AI–powered assistant that can provide answers,

summaries, content generation, and secure task completion based on your enterprise data.

Amazon Q Developer supports developers and IT professionals in various tasks, including

coding, testing, application updates, error diagnosis, security assessments, and AWS resource
optimization.

Features:

Amazon Q offers advanced planning and reasoning abilities to transform and implement new
code features as requested by developers.

Amazon Q is capable of understanding and respecting current governance identities, roles, and
permissions, providing personalized interactions.

Amazon Q integrates with:

● Amazon QuickSight
● Amazon Connect
● AWS Supply Chain

—Back to Index— 31
 Applications of Foundation Models
Prompt Engineering
● Prompts are a specific set of inputs provided by the user that direct the LLMs on
Amazon Bedrock to produce relevant outputs.
● Prompt engineering is the process of designing text prompts to obtain desired
responses from a Large Language Model (LLM). It is the art of communicating with an
LLM.
● Prompt engineers use trial and error methods to generate input texts that guide an
application's generative AI to work as expected.

Benefits of Prompt Engineering:

● Greater developer control
● Improved user experience
● Increased flexibility
Prompt Engineering Techniques:
● Chain-of-thought prompting:- Chain-of-thought prompting involves dividing complex
questions into smaller, logical steps, similar to a thought process. This technique
enhances the model's ability to reason and solve problems effectively.
● Tree-of-thought prompting:- The tree-of-thought technique extends chain-of-thought
prompting by generating multiple potential next steps and evaluating each using a tree
search approach.
● Maieutic prompting:- Maieutic prompting is similar to tree-of-thought prompting. The
model is asked to answer a question with an explanation, followed by prompts to
explain parts of the explanation. Inconsistent explanations are eliminated, improving
performance in complex commonsense reasoning.
● Complexity-based prompting:- This prompt-engineering method utilizes multiple
chain-of-thought rollouts, selecting those with the longest thought sequences and the
most frequent conclusions.
● Generated knowledge prompting:- This technique involves generating relevant facts to
support the prompt, and then completing the prompt. This strategy often results in
higher-quality completions as the model is guided by relevant information.
● Least-to-most prompting:- This method involves prompting the model to identify and
address subproblems sequentially. This allows subsequent subproblems to benefit from
solutions to earlier ones.
● Self-refine prompting:- In this approach, the model is prompted to solve a problem,
evaluate its own solution, and then improve upon it by considering the original problem,
its solution, and its evaluation. This cycle repeats until a predetermined reason to stop.

—Back to Index— 32
 ● Directional-stimulus prompting:- This prompt engineering method employs a hint or
cue, like specified keywords, to direct the language model toward the intended
outcome.
Use-Cases:
Subject matter expertise:
● Example:- Imagine a doctor using a language model to generate potential diagnoses for
a complex patient. By inputting symptoms and patient details, the AI, guided by carefully
crafted prompts, can list possible diseases and narrow down the options based on
additional information.

Critical thinking:
● Example:- In decision-making, a model could be prompted to evaluate various options,
weigh their pros and cons, and suggest the most suitable course of action.

Creativity:
● Example:- Writers can use prompt-engineered models to brainstorm ideas for stories,
generating characters, settings, and plot points. Graphic designers might employ these
models to generate color palettes that convey specific emotions and then create designs
using those palettes.

—Back to Index— 33
 Retrieval Augmented Generation (RAG)
Retrieval Augmented Generation (RAG) is a process to enhance the capabilities of large
language models by referencing an authoritative knowledge base beyond its original training
data, resulting in more accurate and informative responses.
● Large Language Models (LLMs) are trained on massive datasets with billions of
parameters to generate original output for tasks such as answering questions, translating
languages, and completing sentences. Retrieval Augmented Generation (RAG) enhances
the capabilities of LLMs by integrating them with domain-specific or organizational
knowledge bases.
● This approach avoids the need for model retraining and provides a cost-effective way to
ensure that LLM outputs remain accurate, relevant, and valuable in diverse contexts.
Benefits of Retrieval Augmented Generation:
● Cost-effective implementation
● Current information
● Enhanced user trust
● More developer control
How does Retrieval Augmented Generation work?
LLM incorporates the new knowledge and its training data to create better responses.
● Create external data-
○ The new data outside the LLM's original training data set is called external data.
This data can come from various sources like APIs, databases, or document
repositories and may exist in different formats such as files, database records, or
long-form text.
○ Another AI technique, embedding language models, transforms textual data into
numerical representations stored in vector databases, creating a knowledge
library that the generative AI models can understand.
● Retrieve relevant information-
○ The next step is to perform a relevancy search. User queries are also converted
into vectors to retrieve relevant information. The relevancy was calculated and
established using mathematical vector calculations and representations.
● Augment the LLM prompt-
○ The RAG model can generate more accurate and informative responses to user
queries by integrating this retrieved data into the prompts provided by the large
language model. This technique, known as prompt engineering, facilitates
effective communication between the user and the AI.

● Update external data-

—Back to Index— 34
○ To ensure that retrieved information is up-to-date, it's essential to refresh the
documents and update their corresponding embeddings asynchronously. This
can be achieved through automated processes that happen in real time or by
scheduling regular updates.

Figure: Retrieval Augmented Generation

—Back to Index— 35
 RLHF - Reinforcement Learning from Human Feedback
● Reinforcement learning from human feedback (RLHF) is a machine-learning technique
that enhances ML model performance by incorporating human feedback.
● Reinforcement learning (RL) techniques train software to make decisions that maximize
rewards, making their outcomes more accurate.
How does RLHF work?
RLHF involves a four-step process to prepare a model ready for deployment.
● Data collection:- A collection of human-written prompts and their respective answers is
established to serve as training material for the language model.
● Supervised fine-tuning of a language model:- You can use a commercial pre-trained
model as the base model for the RLHF process. You can fine-tune the model to the
company’s internal knowledge base using the retrieval-augmented generation (RAG)
technique. After fine-tuning, evaluate the model's output against predetermined
prompts by comparing its responses to human-generated examples collected in the
initial stage.
● Building a separate reward model:- From a set of multiple responses from the model
answering the single prompt, human evaluators can express their preferences for each option.
By analyzing these ratings, we construct a reward model capable of automatically predicting how
a human would score any given response.
● Optimize the language model with the reward-based model:- The language model uses
the reward model to iteratively refine its response generation strategy. By internally
assessing various potential responses, the model selects the one that it predicts will
yield the highest reward according to the reward model's criteria. This adaptive
approach ensures that the model's outputs align more closely with human preferences.

Figure: RLHF learning process

—Back to Index— 36
Applications of RLHF:
● RLHF can be used in AI image generation: for example, gauging the degree of realism,
technicality, or mood of artwork
● RLHF can generate music that aligns with specific emotions or create soundtracks.
● RLHF can enhance a voice assistant's tone, making it more approachable, curious, and
reliable.
How can AWS services be utilized to fulfill RLHF requirements?
Amazon SageMaker Ground Truth simplifies the task of labeling and annotating data for
Reinforcement Learning from Human Feedback (RLHF), guaranteeing accurate and reliable
human input.
Recall-Oriented Understudy for Gisting Evaluation [ROUGE]:
● ROUGE, or Recall-Oriented Understudy for Gisting Evaluation, is a suite of metrics
commonly used in natural language processing (NLP) to evaluate the quality of
machine-generated text.
● These metrics primarily focus on comparing the generated text and the ground truth
reference (human-written) text of a validation dataset.
● ROUGE measures are designed to assess various aspects of text similarity, such as the
precision and recall of n-grams (contiguous sequences of words) in system-generated
and reference texts.
● The goal is to determine how effectively the model can generate text that is similar to
the original content.

—Back to Index— 37
 Guidelines for Responsible AI
Responsible AI
Responsible AI refers to the development of AI systems that are fair, transparent, accountable,
safe, and unbiased.
Components of responsible AI:
➢ Fairness - Considering the potential effects on diverse groups.
➢ Explainability - Understanding and assessing system outputs.
➢ Privacy and security - Obtaining, utilizing, and securing data and models.
➢ Safety - Safeguarding against system failures and malicious use.
➢ Controllability - Establishing mechanisms to monitor and control AI actions.
➢ Veracity and robustness - Achieving accurate system outputs, under both normal and
adverse conditions.
➢ Governance - Promoting responsible AI by integrating best practices across the AI supply
chain.
➢ Transparency - Ensuring stakeholders have the knowledge required to engage effectively
with the AI system.
Services and tools to build Responsible AI:
Foundation model (FM) evaluations:-
● Model Evaluation on Amazon Bedrock
● Amazon SageMaker Clarify
Implementing safeguards in generative AI:-
● Guardrails in Amazon Bedrock
Detecting bias:-
Biases are imbalances in data or disparities in the performance of a model across different
groups.
● Amazon SageMaker Clarify
Explaining model predictions:-
● Amazon SageMaker Clarify
Monitoring and human review
● Amazon SageMaker Model Monitor
● Amazon Augmented AI
Improving governance
● ML Governance from Amazon SageMaker

—Back to Index— 38
 Amazon SageMaker Clarify
Machine learning offers opportunities to identify and measure bias throughout the ML lifecycle.
Amazon SageMaker Clarify helps detect bias in data and models before, during, and after
training:
1. Pre-Training Bias: Detect bias in the raw data before model training begins.
2. Post-Training Bias: Measure bias in the model’s outputs after training.
3. Monitoring Bias: Continuously monitor bias in model predictions after deployment.
Benefits of SageMaker Clarify:
● Evaluate foundation models (FMs) in minutes:- Automate the evaluation of foundation
models for your generative AI applications based on criteria such as accuracy, resilience,
and bias to uphold responsible AI principles.
● Build trust in ML models:- Assess your FM's performance during customization using
both automated and human-based methods.
● Accessible, science-based metrics and reports:- Generate user-friendly metrics, reports,
and practical examples to support the FM customization and MLOps workflow.
SageMaker Clarify’s Strategy for Addressing Bias
● Bias Metrics: SageMaker Clarify offers model-agnostic metrics to measure bias and
fairness based on different fairness concepts.
● Automation: SageMaker Clarify automates bias detection and monitoring throughout
the ML lifecycle.
● Data Monitoring: SageMaker Clarify tracks bias in model predictions after deployment,
ensuring continuous oversight of model behavior.
● Tools for Bias Detection
● SageMaker Clarify Sample Notebooks: SageMaker Clarify provides a notebook for bias
detection and explainability, helping users run bias detection jobs and interpret feature
attributions.
The sample notebook for bias detection can run in Amazon SageMaker Studio using Python 3
(Data Science). It walks through the process of detecting bias and explaining model predictions.
SageMaker Clarify offers comprehensive tools to measure and monitor bias, helping ensure
machine learning models are fair and unbiased across all development and deployment stages.

—Back to Index— 39
 Amazon SageMaker Model Monitor
Amazon SageMaker Model Monitor monitors the quality of Amazon SageMaker machine
learning models in production.
Model Monitor empowers you to implement continuous monitoring using various approaches;
● Continuous monitoring with a real-time endpoint.
● Continuous monitoring with a batch transform job that runs regularly.
● On-schedule monitoring for asynchronous batch transform jobs.
○ Model Monitor empowers you to establish alerts that notify model quality
deviations. By promptly identifying these deviations, you can take corrective
actions such as retraining models, auditing upstream systems, or fixing quality
issues.
○ Model Monitor offers both pre-built and customizable monitoring options. You
can use the pre-built features for quick setup or write your code for custom
analysis.
Model Monitor offers various kinds of monitoring:
● Monitor data quality - Monitor drift in data quality.
● Monitor model quality - Monitor drift in model quality metrics, such as accuracy.
● Monitor Bias Drift for Models in Production - Monitor bias in your model's predictions.
● Monitor Feature Attribution Drift for Models in Production - Monitor drift in feature
attribution.
Amazon SageMaker Model Monitor continuously monitors the performance of your
machine-learning models in production. It automatically alerts you if any significant changes or
quality issues are detected. Model Monitor uses rules to identify deviations from expected
behavior and notify you promptly.

Figure: Amazon SageMaker Model Monitor

—Back to Index— 40
Amazon SageMaker Model Cards:
● Model governance is a framework that provides systematic visibility into machine
learning (ML) model development, validation, and usage. Amazon SageMaker provides
purpose-built ML governance tools for managing control access, activity tracking, and
reporting across the ML lifecycle.
● Amazon SageMaker Model Cards are essentially a standardized template to document,
retrieve, and share essential model information from development to deployment.
Features:
● Provide guidance on the appropriate use of the model.
● Support audit processes by offering detailed information on model training and
performance metrics.
● Communicate specific business value that the model is expected to deliver.

—Back to Index— 41
 AWS Managed AI Services
Amazon Polly
What is Amazon Polly?
Text-to-Speech Conversion: Transforms written text into spoken words.
Cloud-Based Service: Operates in the cloud, eliminating the need for local infrastructure.
Voice Options: Provides various voice choices, including Neural Text-to-Speech (NTTS) for
natural-sounding speech.
Features
● No Setup Fees: Pay only for the text converted into speech, with no initial setup costs.
● Multilingual Support: Access various languages and Neural Text-to-Speech (NTTS) voices
for creating speech-enabled applications.
● Speech Caching and Replay: Utilize caching and replay features for Amazon Polly’s
generated speech, available in formats like MP3.

Amazon Comprehend
What is Amazon Comprehend?
Natural Language Processing (NLP): Uses NLP to extract insights from document content.
Insight Extraction: Identifies entities, key phrases, language, sentiments, and other elements
within documents.
Product Development: Leverage document structure understanding to develop new products.
Features:
● Extract Insights from Diverse Text Sources: Analyze text from documents, support
tickets, product reviews, emails, and social media to uncover valuable insights.
● Optimize Document Processing: Improve workflows by extracting key information,
including text, phrases, topics, and sentiment from documents such as insurance claims.
● Custom Document Classification: Differentiate your business by training models to
classify documents and recognize specific terms, without needing advanced machine
learning skills.
● Protect Sensitive Information: Safeguard and manage access to sensitive data by
identifying and redacting Personally Identifiable Information (PII) in documents.
Use cases:
● Analyze business and call center data
● Index and search through product reviews
● Manage legal briefs
● Handle financial document processing

—Back to Index— 42
 Amazon Rekognition
What is Amazon Rekognition?
Cloud-Based Service: Utilizes advanced computer vision technology for image and video
analysis.
No Machine Learning Expertise Needed: Accessible through an intuitive API.
Integration with Amazon S3: Quickly analyzes images and videos stored in Amazon S3.
Key Features:
Includes object and text detection, unsafe content identification, and facial analysis.
Facial analysis
● User Verification: Identifies and verifies individual identities.
● Cataloging: Organizes and manages face data for various applications.
● Public Safety: Enhances safety through surveillance and monitoring.
● Detect, analyze, and compare faces in both live streaming and recorded videos.
Image Analysis:
● Object, Scene, and Concept Detection: Detect and classify various objects, scenes,
concepts, and celebrities present in images.
● Text Detection: Identify both printed and handwritten text in images, supporting
multiple languages.
Video Analysis:
● Object, Scene, and Concept Detection: Categorize objects, scenes, concepts, and
celebrities appearing in videos.
● Text Detection: Recognize printed and handwritten text in videos in different languages.
● People Tracking: Monitor individuals identified in videos as they move across frames.
Use cases:
● Simplify content retrieval with Amazon Rekognition's automatic analysis, enabling easy
searchability for images and videos.
● Enhance security with Rekognition's face liveness detection, preventing identity spoofing
beyond traditional passwords.
● Quickly locate individuals across your visual content using Rekognition's efficient face
search feature.
● Ensure content safety with Rekognition's ability to detect explicit, inappropriate, and
violent content, facilitating proactive filtering.
● Benefit from HIPAA Eligibility, making Amazon Rekognition suitable for handling
protected health information in healthcare applications.

—Back to Index— 43
 Amazon Lex
What Is Amazon Lex?
Build Chatbots: Create conversational interfaces using natural language processing.
Leverages Alexa Technology: Utilizes the same technology that powers Alexa for advanced
language understanding.
Seamless Integration: Easily integrates with other AWS services to enhance chatbot
functionality and user experience.
Features:
● Effortlessly integrate AI that comprehends intent, retains context, and automates basic
tasks across multiple languages.
● Design and deploy omnichannel conversational AI with a single click, without the need
to manage hardware or infrastructure.
● Seamlessly connect with other AWS services to access data, execute business logic,
monitor performance, and more.

● Pay only for speech and text requests without any upfront costs or minimum fees.
Use Cases:
● Enable virtual agents and voice assistants: Provide users with self-service options
through virtual contact center agents and interactive voice response (IVR), allowing
them to perform tasks autonomously, like scheduling appointments or changing
passwords.
● Automate responses to FAQs: Develop conversational solutions that answer common
inquiries, enhancing Connect & Lex conversation flows with natural language search
for frequently asked questions powered by Amazon Kendra.
● Improve productivity with application bots: Streamline user tasks within applications
using efficient chatbots, seamlessly integrating with enterprise software through AWS
Lambda and maintaining precise access control via IAM.
● Extract insights from transcripts: Design chatbots using contact center transcripts to
maximize captured information, reducing design time and expediting bot deployment
from weeks to hours.

—Back to Index— 44
 Amazon Transcribe
What is Amazon Transcribe?
Speech-to-Text Conversion: Transforms audio speech into written text.
Deep Learning Technology: Utilizes automatic speech recognition (ASR) for accurate
transcription.
Versatile Applications: Ideal for generating transcripts from various audio sources, such as
meetings, interviews, and videos.
Features
● Optimized for Specific Use Cases: Ideal for customer service calls, live broadcasts, and
media subtitling.
● Medical Transcription: Converts medical speech to text for clinical documentation with
high accuracy.
● Cost Structure: Charges are based on the seconds of speech converted per month.
Use Cases
● Customer Service: Enhance customer interactions by transcribing service calls for
analysis and improvement.
● Live Broadcasts: Generate real-time subtitles for live events and broadcasts.
● Medical Documentation: Streamline clinical documentation by transcribing medical
speech accurately.

—Back to Index— 45
 Amazon Translate
What is Amazon Translate?
Neural Machine Translation: Uses neural networks for accurate and natural text translations.
Language Pairs: Translates text between English and multiple other languages.
Source-Target Conversion: Converts text from a source language to a target language based on
selected language pairs.
Benefits of Amazon Translate:
● High-quality translations - Provide precise and evolving translations across various
applications.
● Batch and real-time translations - Integrate batch and real-time translation into your
applications seamlessly using a single API call.
● Customization - Customize your ML–translated output to define brand names, model
names, and other unique terms.
Use cases:
● Translate user-generated content: Automatically translate user-generated content,
including social media posts, profiles, and comments, instantly in real-time.
● Analyze online conversations in different languages: Use a natural language processing
application to analyze text in multiple languages and gain insights into public opinion
about your brand, product, or service.
● Create cross-lingual communications between users: Implement real-time language
translation capabilities in chat, email, helpdesk, and ticketing systems to enable
English-speaking agents to communicate effectively with customers worldwide.

—Back to Index— 46
 Amazon Mechanical Turk (MTurk)
What is Amazon Mechanical Turk?
● Crowdsourcing Marketplace: MTurk connects individuals and businesses with a global,
virtual workforce for various tasks.
● Task Types: Includes simple data validation, research, survey participation, content
moderation, and more.
● How It Works: Requesters post tasks (HITs) that Workers complete online. The system
ensures workers are paid only for satisfactory work, and you can use qualification tests
to select skilled Workers.
Advantages
● Enhanced Efficiency: Automate repetitive, manual tasks to streamline workflows. MTurk
helps complete tasks quickly, freeing up internal resources for more strategic work.
● Flexible Scaling: Easily scale workforce up or down without the complexities of
managing a temporary in-house team. MTurk provides access to a 24x7 global
workforce.
● Cost Reduction: Lower labor and overhead costs with a pay-per-task model. MTurk helps
manage expenses effectively while achieving results that might be challenging with a
dedicated team.
Why Use MTurk?
● Human Expertise: Completes tasks better than computers, such as content moderation
and data deduplication.
● Efficient Crowdsourcing: Breaks down complex projects into manageable microtasks for
distributed workers, improving scalability and reducing manual effort.
MTurk Use Cases in Machine Learning
Data Collection and Annotation
● Efficient Data Gathering: MTurk simplifies the collection and labeling of large datasets
needed for training ML models. It accelerates the process of annotating data, such as
tagging images or categorizing text.
Model Improvement
● Continuous Iteration: Use MTurk for ongoing adjustments and enhancements to ML
models. Human input helps in refining models by providing feedback and making
necessary corrections.
Human-in-the-Loop (HITL)
● Incorporating Human Feedback: MTurk supports HITL workflows where human
feedback is essential for model validation and retraining. For instance, annotating
images with bounding boxes helps create precise datasets for computer vision tasks,
especially when automated solutions fall short.

—Back to Index— 47
 Amazon Augmented AI [Amazon A2I]
● Amazon Augmented AI (Amazon A2I) is a service that brings human review of ML
predictions to all developers by removing the heavy lifting associated with building
human review systems or managing large numbers of human reviewers.
● For example, extracting information from scanned mortgage application forms can
require human review due to low-quality scans or poor handwriting.
● Building human review systems can be time-consuming and expensive because it
involves implementing complex workflows, writing custom software to manage review
tasks and results, and managing large groups of reviewers.
Amazon A2I offers built-in human review processes for typical ML tasks like content moderation
and document text extraction.
● Content moderation
● Form extraction
● Image classification
Use cases:
➔ Use Amazon A2I with Amazon Textract:- Use Amazon Textract to randomly select and
send documents from your dataset to humans for review.
➔ Use Amazon A2I with Amazon Rekognition:- Use Amazon Textract to randomly select
and send images from your dataset to humans for review.
➔ Use Amazon A2I to review real-time ML inferences:- Use Amazon A2I to review
real-time, low-confidence inferences made by a model deployed to a SageMaker.
Continuously refine your model by incorporating feedback from A2I's output data.
➔ Use Amazon A2I with Amazon Comprehend:- Have humans to review inferences about
text data using Amazon Comprehend.
➔ Use Amazon A2I with Amazon Transcribe:- Have humans to review transcriptions of
video or audio files using Amazon Transcribe.

—Back to Index— 48
 AWS DeepRacer
AWS DeepRacer is a reinforcement learning (RL)-enabled autonomous 1/18th-scale vehicle,
designed for educational purposes with supporting services in the AWS Machine Learning
ecosystem.
● AWS DeepRacer offers a platform for developing and testing deep reinforcement
learning models. You can train these models in a simulated environment and then
deploy them to a physical AWS DeepRacer car for autonomous racing.
AWS DeepRacer vehicles:
1. The original AWS DeepRacer device is a physical 1/18th-scale model car with a mounted
camera and an onboard compute module. The compute module runs inference in order
to drive itself along a track. The compute module and the vehicle chassis are powered by
dedicated batteries known as the compute battery and the drive battery, respectively.
2. The virtual race car becomes the original AWS DeepRacer device, the Evo device, or
various digital rewards that can be earned by participating in AWS DeepRacer League
Virtual Circuit races.
3. The AWS DeepRacer Evo device is the original device with an optional sensor kit. The kit
includes an additional camera and LIDAR (light detection and ranging), that allow the car
to detect objects behind and lateral to itself.

—Back to Index— 49
 Security, Compliance, and Governance for AI Solutions
Amazon Macie
What is Amazon Macie?
● Amazon Macie is a data security solution that employs machine learning algorithms and
pattern recognition techniques to identify and safeguard sensitive data.
● By leveraging machine learning and pattern-matching capabilities, Amazon Macie not
only detects sensitive data but also offers insights into potential data security threats.
● Additionally, it facilitates automated measures to mitigate these risks, enhancing overall
data protection.
Features:
● Implement automated processes for detecting sensitive data on a large scale.
● Use Amazon Macie with Amazon Textract, Amazon Rekognition, and Amazon SageMaker
to discover and secure sensitive data stored in Amazon S3.

AWS PrivateLink
What is AWS PrivateLink?
● AWS PrivateLink is a network service used to connect to AWS services hosted by other
AWS accounts (referred to as endpoint services) or AWS Marketplace.
● Whenever an interface VPC endpoint (interface endpoint) is created for service in the
VPC, an Elastic Network Interface (ENI) in the required subnet with a private IP address is
also created that serves as an entry point for traffic destined to the service.
Interface endpoints
● It serves as an entry point for traffic destined to an AWS service or a VPC endpoint
service. Gateway endpoints
● It is a gateway in the route-table that routes traffic only to Amazon S3 and DynamoDB.
Features:
● It provides security by not allowing the public internet and reducing the exposure to
threats, such as brute force and DDoS attacks.
● With the help of AWS PrivateLink, VPC interface endpoint connects your VPC directly to
the SageMaker API or SageMaker Runtime without using an internet gateway, NAT
device, VPN connection.
● Define an interface VPC endpoint for Amazon Rekognition to connect your VPC to
Amazon Rekognition.

—Back to Index— 50