AWS AI Exam Guide

• The AWS AI exam includes these content domains and weightings:
  • Fundamentals of AI and ML: 20%
  • Fundamentals of Generative AI: 24%
  • Applications of Foundation Models: 28%
  • Guidelines for Responsible AI: 14%
  • Security, Compliance, and Governance for AI Solutions: 14%

Fundamentals of AI and ML

• Artificial intelligence involves creating intelligent systems that perform tasks requiring human intelligence

• Machine learning enables machines to learn by understanding and building methods

• Deep learning uses neurons and synapses similar to the human brain for learning

• Generative AI, a deep learning subset, adapts pre-built models without retraining or fine-tuning and can generate new data from learned patterns

Machine Learning Fundamentals

• Training data involves collecting and preparing data, selecting an algorithm, training the model, then evaluating performance through testing and iteration
• Labeled data includes instances with labels representing the desired output, such as images labeled with corresponding class labels
• Unlabeled data lacks labels and target variables, consisting only of input features and lacking corresponding output or classes

Data Structures

• Structured data is organized in a predefined manner, such as tables or databases, and is suitable for traditional ML algorithms

• Tubular data is stored in spreadsheets or CSV files with instances as rows and features as columns

• Time-series data consists of sequential values measured over time, like stock prices or weather data

• Unstructured data lacks a predefined format like text, images, audio, or video, requiring advanced ML techniques

Machine Learning Process

• The ML process involves feeding training data into algorithms, categorized into supervised, unsupervised, and reinforcement learning

Supervised Learning

• Supervised learning algorithms train on labeled data to learn a mapping function, predicting outputs for new, unseen data

Unsupervised Learning

• Unsupervised learning uses algorithms that learn from unlabeled data to discover inherent patterns or relationships

Reinforcement Learning

• Reinforcement learning provides a performance score and semi-supervised learning, with partial labeled data
• Feedback is given as rewards or penalties, allowing the machine to learn from its actions

Inferencing

• Inferencing begins after model training, using learned information to make predictions or decisions, with batch and real-time inferencing as main types

Batch Inferencing

• Batch inferencing processes large datasets all at once for tasks where accuracy is more important than speed, like data analysis

Real-time Inferencing

• Real-time inferencing requires quick decision-making in response to new information, crucial in applications like chatbots

Deep Learning Fundamentals

• Deep learning is inspired by brain structure, using artificial neural networks to mimic human information processing
• Neural networks are structured in layers, with interconnected nodes that can be input, hidden or output layers

Neural Networks

• Neural networks identify patterns by adjusting connections between nodes, like identifying customer types based on purchasing behavior

AI Branches Using Deep Learning

• Computer vision enables computers to interpret digital images and videos
• Natural Language Processing (NLP) manages the interaction between computers and human languages

Generative AI Fundamentals

• Foundation Models (FMs) are pre-trained on internet-scale data.

FM Lifecycle

• Data Selection: unlabeled data can be used at scale because it is easier to obtain than labeled, and needed to train require massive datasets from diverse sources
• Pre-training: FMs are pre-trained though self-supervised learning without labelled data. Algorithms may identify relationships in the data such as the relationship between the word "drink" and beverage
• Optimization: FM's can be optimized through prompt engineering, Retrieval-Augmented Generation (RAG) and fine-tuning of task-specific data
• Evaluation: FM's performance can be measured using appropriate metrics and benchmarks to identify any business needs are met
• Deployment: Once the FM's performance has been confirmed it can be deployed into the target production environment, and integrated into any application through APIs etc
• Feedback and continuous improvement: The models performance is continuously monitored and feedback is collected from users and domain experts

Large Language Models

• Transformer-based LLMs are models that can understand and generate human-like text. They are trained on text data from the internet and learn relationships between words

Tokens

• Tokens are the basic units of text such as words and phrases to standardize input data

Embeddings and Vectors

• Embeddings are numerical representations of tokens which are assigned a vector that captures the token's meaning within the model and which are learned throughout the training process

Diffusion Models

• Forward diffusion gradually adds small amounts of noise to an image until only noise is left
• Reverse diffusion gradually denoises the noisy image until a new image is generated

Multimodal Models

• Multimodal models process and generate multiple data modes like text and images simultaneously, learning how different modalities connect
• Multimodal models automate tasks like video captioning, creating graphics, smarter question answering and translating content while keeping visuals

Generative Adversarial Networks (GANs)

• GANs involve two neural networks, a generator and discriminator, in a zero-sum game where the Generator creates synthetic data from random noise, resembling training data
• The Discriminator distinguishes real data from synthetic data and learns what's what

Variational Autoencoders (VAEs)

• VAEs combine autoencoders and variational inference which consist of two parts
• Encoder transforms input data into a lower-dimensional latent space
• Decoder transforms the latent representation from the encoder, and generates a reconstruction of the data

Optimizing Model Outputs

• Prompt engineering involves creating instructions for foundation models through developing, designing and optimising prompts to enhance output

Prompt Engineering

• Prompts guide the model’s behavior and include:
  • Instructions: Providing task descriptions
  • Context: External guiding information
  • Input Data: The input which for you want a response
  • Output indicator: output type and format

Fine-tuning

Fine-tuning is a supervised learning process using smaller datasets to narrow a pre-trained model's focus

• Instruction fine-tuning uses labeled examples on how the model should respond
• Reinforcement Learning from Human Feedback (RLHF) uses human feedback data to align model output to align with human preferences

Retrieval-Augmented Generation

• Retrieval-augmented generation supplies domain-relevant data as context for responses, using relevant documents to answer user prompts
• RAG does not change the foundation model's weights like fine-tuning does

AWS Infrastructure and Technology

• Amazon SageMaker provides the tool components for ML in a single toolset, making models get to production faster for less effort and cost

AI/ML Services

• Amazon Comprehend uses ML and Natural language processing (NLP) discovering relationships and insights in unstructured data, and includes things like identifying the language of the text
• Amazon Translate neural machine translation delivers fast, high-quality, and affordable language translation, localising content and efficiently implementing cross-lingual communication
• Amazon Textract automatically extracts text and data from scanned documents, identifying form fields and information
• Amazon Lex is a fully managed AI service that has advanced deep learning functionalities of automatic speech recognition (ASR) for converting speech to text, and natural language understanding (NLU) to recognize the intent of the text.

Amazon Services

• Amazon Polly turns given text into lifelike speech using AI that sounds like a human voice
• Amazon Transcribe transcribes the converting of speech to text, including voice-based customer service calls
• Amazon Rekognition facilitates adding image and video analysis to your applications by identifying text, scenes, activities and any other objects, people within as well as facial analysis
• Amazon Kendra is an intelligent search service that is powered by Machine Learning
• Amazon Personalize creates individualized recommendations for customers.

AWS DeepRacer

• AWS DeepRacer is the 1/18th scale race car that provides a fun way to get started with reinforcement learning (RL) due to it being a different approach to training models with advanced machine learning techniques

Generative AI

• Amazon SageMaker JumpStart helps quickly get started with Machine Learning and is deployed to any set of solutions for the most used use cases
• Amazon Bedrock is a fully managed service that makes Foundation Models from Amazon through any API using a serverless experience
• Amazon Q helps get fast, relevant answers to questions and generate any actions and content using an expertise in a companys information
• Amazon Q developer provides Machine Learning powered code recommendations through things such as c#, Python Java etc

Advantages/Benefits of AWS AI Solutions

• Accelerated development and deployment is achieved by Amazon Q Developer (previously Amazon CodeWhisperer) which can generate code, SageMaker which helps such as data processing, Bedrock which provides access to pre-trained models and APIs
• Scalability and Cost Optimization is same as AWS cloud
• Flexibility and access to models as AWS is continuously updating and expanding its AI services and providing access to latest advancements in machine learning models and algorithms
• Integration with AWS tools and services which offers readily AI ready capabilities that can be implemented into applications

Cost Considerations

• Responsiveness and availability
• Redundancy and regional coverage
• Availability
• Performance
• Token-based pricing
• Provisioned throughput
• Custom models