AWS Certified AI Practitioner

AI

Artificial Intelligence is the general development of intelligent systems capable of performing tasks that typically require human intelligence,

ML

Machine Learning is a type of AI for understanding and building methods that make it possible for machines to learn

DL

Deep Learning uses the concept of neurons and synapses similar to how our brain is wired. E.g. Amazon Rekognition

Generative AI

A subset of deep learning because it can adapt models built using deep learning, but without retraining or fine tuning

Machine Learning Process

Training data -> ML Algorithm -> Model

Types of ML data used in training an ML model

Labeled, unlabeled, structured, and unstructured data

Labeled vs Unlabeled data

Each instance is accompained by a label or target variable representing, vs no associated label

Structured vs Unstructured data

Data is organised in a predefined manner (csv, database, time, readings) vs lacking any format (text, images, audio, and video)

Supervised learning

Algorithms trained on labeled data, to learn a mapping function to predict the output for a new data input

Unsupervised learning

Algorithms learn from unlabeled data, to discover inherent patterns, structures, or relationships within the input data

Reinforcement learning

The machine is given a score as guidance. Semi supervised learning. Learns from feedback to in form of reward or penalty

Batch inferencing

Computer takes a large amount of data and analyses it all at once to provide a set of results. For data analsis or accuracy is not crucial

Real-time inferencing

Computer must make decisions as they come in, decision making is critical and instantaneous (E.g. Self driving cars or chatbots)

Neural Networks

Contains nodes in layers, Identifies patterns from examples, make predictions on new data

NLP

Branch of AI dealing with computer and human language interactions

Computer Vision

Computers can interpret and understand unstructured data

FMs

Foundation Models, pretrained on internet-scale data, powers gen ai

FM Lifecycle

Data selection -> Pre-training -> Optimization -> Evaluation -> Deployment -> Feedback and Continuous Improvement

How to optimize FMs

Prompt engineering, fine-tuning, or RAG

LLMs

Large Language Models,

Powerful models that can understand and generate human-like text. Trained on internet, books and more

Tokens

Basic units of text that the model processes. Standardize input data, making easy processing

Embeddings

Numerical reps of tokens, assigning to a vector (list), capturing meaning and relationship with other tokens

Diffusion models

DL architecture system starting with pure noise or random data then gradually adding more meaningful information

Forward vs Reverse Diffusion

Adding small amount of noise until only noise vs noisy image is denoised until a image is introduced

Multimodal models

Process and general multiple modes of data simultaneously (e.g. image + text = new image + caption, auto video captioning)

GAN

Generative Adversarial Networks

Gen model involving 2 neural networks competing (generator (synthetic) and discriminator (real))

VAE

Variational Autoencoders

Combines autoencoders and Variational inference. Contains encoder (input) and decoder (reconstruct)

Prompt engineering

Prompts act as instructions for foundation model. Focuses on developing, designing, and optimizing prompts to enhance the output

Prompt forms

Instructions, Context, Input data, Output indicator

Fine-tuning

Supervised learning process, adding specific, smaller datasets. Instruction or RLHF

Changes the model weights

Instruction vs RLHF fine-tuning

Using examples (prompt tuning) compared to reinforcement learning from human feedback (data)

RAG

retrieval-augmented generation

Supplies domain-relevant data as context to product responses based on that data.

Does not change the weights of the FM

Amazon Sagemaker

Fully managed ML to build, train, and deploy custom models

Amazon Comprehend

Natural Language Processing (NLP)

Amazon Translate

Language translation

Amazon Textract

Extracting data from scanned documents

Amazon Lex

Build conversational interfaces (Same DL as Alexa)

Amazon Polly

Text to speech

Amazon Transcribe

Automatic speech recognition

Amazon Rekognition

DL based analysing images and videos

Amazon Kendra

Enterprise search for websites and applications to readily find content

Amazon Personalize

Real time personalization and recommendations

AWS DeepRacer

Fully autonomous 1/18th scale race car for hands on experience with reinforcement learning

Amazon SageMaker JumpStart

Provides a set of solutions for the most common use cases

Amazon Bedrock

Fully managed service that makes FMs from Amazon and leading AI startups through an API

PartyRock

Hands-on experience experiementing with building AI applications

Amazon Q

A gen AI powered assistant designed for work that can be tailored for a business's data

Amazon Q Developer

Provides ML powered code recommendations to accelerate development in a variety of programming languages and applications

4 Benefits of AWS AI solutions

1. Accelerated development and deployment

2. Scalability and cost optimization

3. Flexibility and access to models

4. Integration with AWS tools and services

Cost considerations

1. Responsiveness and availability

2. Redundancy and regional coverage

3. Performance

4. Token-based pricing

5. Provisioned throughput

6. Custom models

Media and entertainment use cases

Content generation, VR, new generation (articles or summaries)

Retail use cases

Product review summaries, pricing optimization, virtual try-ons, store layout

Healthcare use cases

Personalize medicine, improve medical imaging

AWS HealthScribe

Empowers healthcare software vendors to build clinical applications

Life sciences use cases

Drug discovery, protein folding prediction, synthetic biology

Financial services use cases

Fraud detection mechanisms, portfolio management, debt collection

Manufacturing use cases

Predictive maintenance, process optimization, product design, material science

Computer Vision applications

Autonomous driving, healthcare or medical imaging, and public safety and home security

NLP applications

Insurance (e.g. extract policy no.s, expiration date), telecommunication (e.g. customer messages) , and education (Q&A chatbots)

IDP

Intelligent Document Processing

Extracts and classifies information from unstructured data, generating summaries, and providing actionable insights

IDP applications

Financial services, lending, legal, and healthcare

Fraud detection applications

Used by financial services, retail, and telecommunication

Types of supervised learning

Classification (Assign labels and categories to new data) and regression (Predicting continuous or numerical values)

Types of unsupervised learning

Clustering (Groups data based on similar features or distances) and Dimensionality reduction (reduce the amount of features or dimensions in a dataset)

Capabilities of Gen AI

Adaptability, responsiveness, simplicity, creativity and exploration, data efficiency, personalization, and scalability

Challenges of Gen AI

Regulatory violations, social risks, data security and privacy concerns, toxicity, hallucinations, interpretability, non determinism

Factors to Consider When Selecting a Generative AI Model

Performance requirements, constraints, capabilities, compliance, cost

Gen AI models

Jurassic-2, Amazon Titan, Claude, Stable Diffusion, Command, Llama

Responsible AI

An AI system practicing trustworthiness and transparency

AI bias

the model is missing important features of the datasets

AI variance

the model's sensitivity to fluctuations or noise in the training data

Overfitted

model performs well on the training data but does not perform well on the evaluation data

Underfitted

High bias, the model is not capturing enough difference in the features of the data

How to overcome bias and variance errors

Cross validation, increase data, regularization, simpler models, dimension reduction, stop training early

Responsible AI dimensions

1. Fairness

2. Explainability

3. Privacy and security

4. Veracity and robustness

5. Transparency

6. Governance

7. Safety

8. Controllability

Model Evaluation on Amazon Bedrock

Evaluate, compare, and select the best foundation model for your use case

Amazon SageMaker AI Clarify

Auto evaluate FMs for gen ai use cases with accuracy, robustness, and toxicity

Identifies potential bias in ML

Guardrails for Amazon Bedrock

Help control the interaction between users and FMs by filtering undesirable and harmful content, redacting PII

Amazon SageMaker Data Wrangler

Prepares and more data using web interface. Balances data through random undersampling, random oversampling, and synthetic minority oversampling Technique (SMOTE)

Amazon SageMaker Model Monitor

Monitors the quality of SageMaker ML models in production

Amazon A2I

Amazon Augmented AI

Builds the workflows required for human review of ML predictions and help ensure accurate predictions

Data curation

Curating datasets and ensuring it is relevant, high quality, and reliable data

Data Processing -> Data Augmentation -> Regular Auditing

AWS AI Service Cards

A resource to increase transparency and help customers better understand AWS AI services

SageMaker Model Cards

Document critical details about ML models for streamlined governance and reporting

SageMaker Autopilot

Uses tools provided by SageMaker AI Clarify to help provide insights into how ML models make predictions

Model Trade-Offs

Interpretability, Explainability, Safety, Controllability,

Principles of human-centered design for explainable AI

Design for amplified and unbiased decision-making, human and AI learning, reinforcement learning from human feedback

Amazon SageMaker Ground Truth

Incorporating human feedback across the ML lifecycle to improve model accuracy and relevancy

ML Development Lifecycle

Business Problem → ML Problem Framing → Data Processing → Model Development → Model Deployment → Model Monitoring → Model Retraining

Machine Learning Advanced Development Lifecycle

What happens during the Data Processing stage

Data collection, data pre-processing, and feature engineering

Amazon SageMaker Studio Classic

Built-in integration of Amazon EMR and AWS Glue sessions to handle data prep and ml workflows

Amazon SageMaker Processing API

Customers can run scripts and notebooks to process, transform, and analyse datasets

Amazon SageMaker Feature Store

Helps to create, share, and manage features for ML development. Features can be retrieved and enriched

Amazon SageMaker Canvas

Use machine learning to generate predictions without needing to write any code

Amazon SageMaker Experiments

Experiment with multiple combinations of data, algorithms, and parameters, all while observing the impact of incremental changes on model accuracy.

Amazon SageMaker Automatic Model Tuning

Uses hyperparameter tuning by running many jobs with different combinations and measuring each of them

Hyperparameter

A way to find the best version of your models

Confusion Matrix

Help classify why and how a model gets something wrong

Model Accuracy Calculation

( TP + TN ) / (TP + TN + FP + FN )