The Machine Learning Solutions Architect Handbook: (Final),2nd Edition

Preface
   Who this book is for
   What this book covers
   To get the most out of this book
   Get in touch
Navigating the ML Lifecycle with ML Solutions Architecture
   ML versus traditional software
   ML lifecycle
      Business problem understanding and ML problem framing
      Data understanding and data preparation
      Model training and evaluation
      Model deployment
      Model monitoring
      Business metric tracking
   ML challenges
   ML solutions architecture
      Business understanding and ML transformation
      Identification and verification of ML techniques
      System architecture design and implementation
      ML platform workflow automation
      Security and compliance
   Summary
Exploring ML Business Use Cases
   ML use cases in financial services
      Capital market front office
         Sales trading and research
         Investment banking
         Wealth management
      Capital market back office operations
         Net Asset Value review
         Post-trade settlement failure prediction
      Risk management and fraud
         Anti-money laundering
         Trade surveillance
         Credit risk
      Insurance
         Insurance underwriting
         Insurance claim management
   ML use cases in media and entertainment
      Content development and production
      Content management and discovery
      Content distribution and customer engagement
   ML use cases in healthcare and life sciences
      Medical imaging analysis
      Drug discovery
      Healthcare data management
   ML use cases in manufacturing
      Engineering and product design
      Manufacturing operations – product quality and yield
      Manufacturing operations – machine maintenance
   ML use cases in retail
      Product search and discovery
      Targeted marketing
      Sentiment analysis
      Product demand forecasting
   ML use cases in the automotive industry
      Autonomous vehicles
         Perception and localization
         Decision and planning
         Control
      Advanced driver assistance systems (ADAS)
   Summary
Exploring ML Algorithms
   Technical requirements
   How machines learn
   Overview of ML algorithms
      Consideration for choosing ML algorithms
      Algorithms for classification and regression problems
         Linear regression algorithms
         Logistic regression algorithms
         Decision tree algorithms
         Random forest algorithm
         Gradient boosting machine and XGBoost algorithms
         K-nearest neighbor algorithm
         Multi-layer perceptron (MLP) networks
      Algorithms for clustering
      Algorithms for time series analysis
         ARIMA algorithm
         DeepAR algorithm
      Algorithms for recommendation
         Collaborative filtering algorithm
         Multi-armed bandit/contextual bandit algorithm
      Algorithms for computer vision problems
         Convolutional neural networks
         ResNet
      Algorithms for natural language processing (NLP) problems
         Word2Vec
         BERT
      Generative AI algorithms
         Generative adversarial network
         Generative pre-trained transformer (GPT)
         Large Language Model
         Diffusion model
   Hands-on exercise
      Problem statement
      Dataset description
      Setting up a Jupyter Notebook environment
      Running the exercise
   Summary
Data Management for ML
   Technical requirements
   Data management considerations for ML
   Data management architecture for ML
      Data storage and management
         AWS Lake Formation
      Data ingestion
         Kinesis Firehose
         AWS Glue
         AWS Lambda
      Data cataloging
         AWS Glue Data Catalog
         Custom data catalog solution
      Data processing
      ML data versioning
         S3 partitions
         Versioned S3 buckets
         Purpose-built data version tools
      ML feature stores
      Data serving for client consumption
         Consumption via API
         Consumption via data copy
      Special databases for ML
         Vector databases
         Graph databases
      Data pipelines
      Authentication and authorization
      Data governance
         Data lineage
         Other data governance measures
   Hands-on exercise – data management for ML
      Creating a data lake using Lake Formation
      Creating a data ingestion pipeline
      Creating a Glue Data Catalog
      Discovering and querying data in the data lake
      Creating an Amazon Glue ETL job to process data for ML
      Building a data pipeline using Glue workflows
   Summary
Exploring Open-Source ML Libraries
   Technical requirements
   Core features of open-source ML libraries
   Understanding the scikit-learn ML library
      Installing scikit-learn
      Core components of scikit-learn
   Understanding the Apache Spark ML library
      Installing Spark ML
      Core components of the Spark ML library
   Understanding the TensorFlow deep learning library
      Installing TensorFlow
      Core components of TensorFlow
      Hands-on exercise – training a TensorFlow model
   Understanding the PyTorch deep learning library
      Installing PyTorch
      Core components of PyTorch
      Hands-on exercise – building and training a PyTorch model
   How to choose between TensorFlow and PyTorch
   Summary
Kubernetes Container Orchestration Infrastructure Management
   Technical requirements
   Introduction to containers
   Overview of Kubernetes and its core concepts
      Namespaces
      Pods
      Deployment
      Kubernetes Job
      Kubernetes custom resources and operators
      Services
   Networking on Kubernetes
   Security and access management
      API authentication and authorization
   Hands-on – creating a Kubernetes infrastructure on AWS
      Problem statement
      Lab instruction
   Summary
Open-Source ML Platforms
   Core components of an ML platform
   Open-source technologies for building ML platforms
      Implementing a data science environment
      Building a model training environment
      Registering models with a model registry
      Serving models using model serving services
         The Gunicorn and Flask inference engine
         The TensorFlow Serving framework
         The TorchServe serving framework
         KFServing framework
         Seldon Core
         Triton Inference Server
      Monitoring models in production
      Managing ML features
      Automating ML pipeline workflows
         Apache Airflow
         Kubeflow Pipelines
   Designing an end-to-end ML platform
      ML platform-based strategy
      ML component-based strategy
   Summary
Building a Data Science Environment Using AWS ML Services
   Technical requirements
   SageMaker overview
   Data science environment architecture using SageMaker
      Onboarding SageMaker users
      Launching Studio applications
      Preparing data
      Preparing data interactively with SageMaker Data Wrangler
      Preparing data at scale interactively
      Processing data as separate jobs
      Creating, storing, and sharing features
      Training ML models
      Tuning ML models
      Deploying ML models for testing
   Best practices for building a data science environment
   Hands-on exercise – building a data science environment using AWS services
      Problem statement
      Dataset description
      Lab instructions
         Setting up SageMaker Studio
         Launching a JupyterLab notebook
         Training the BERT model in the Jupyter notebook
         Training the BERT model with the SageMaker Training service
         Deploying the model
         Building ML models with SageMaker Canvas
   Summary
Designing an Enterprise ML Architecture with AWS ML Services
   Technical requirements
   Key considerations for ML platforms
      The personas of ML platforms and their requirements
         ML platform builders
         Platform users and operators
      Common workflow of an ML initiative
      Platform requirements for the different personas
   Key requirements for an enterprise ML platform
   Enterprise ML architecture pattern overview
      Model training environment
         Model training engine using SageMaker
         Automation support
         Model training lifecycle management
      Model hosting environment
         Inference engines
         Authentication and security control
         Monitoring and logging
   Adopting MLOps for ML workflows
      Components of the MLOps architecture
      Monitoring and logging
         Model training monitoring
         Model endpoint monitoring
         ML pipeline monitoring
         Service provisioning management
   Best practices in building and operating an ML platform
      ML platform project execution best practices
      ML platform design and implementation best practices
      Platform use and operations best practices
   Summary
Advanced ML Engineering
   Technical requirements
   Training large-scale models with distributed training
      Distributed model training using data parallelism
         Parameter server overview
         AllReduce overview
      Distributed model training using model parallelism
         Naïve model parallelism overview
         Tensor parallelism/tensor slicing overview
         Implementing model-parallel training
   Achieving low-latency model inference
      How model inference works and opportunities for optimization
      Hardware acceleration
         Central processing units (CPUs)
         Graphics processing units (GPUs)
         Application-specific integrated circuit
      Model optimization
         Quantization
         Pruning (also known as sparsity)
      Graph and operator optimization
         Graph optimization
         Operator optimization
      Model compilers
         TensorFlow XLA
         PyTorch Glow
         Apache TVM
         Amazon SageMaker Neo
      Inference engine optimization
         Inference batching
         Enabling parallel serving sessions
         Picking a communication protocol
      Inference in large language models
         Text Generation Inference (TGI)
         DeepSpeed-Inference
         FastTransformer
   Hands-on lab – running distributed model training with PyTorch
      Problem statement
      Dataset description
      Modifying the training script
      Modifying and running the launcher notebook
   Summary
Building ML Solutions with AWS AI Services
   Technical requirements
   What are AI services?
   Overview of AWS AI services
      Amazon Comprehend
      Amazon Textract
      Amazon Rekognition
      Amazon Transcribe
      Amazon Personalize
      Amazon Lex V2
      Amazon Kendra
      Amazon Q
      Evaluating AWS AI services for ML use cases
   Building intelligent solutions with AI services
      Automating loan document verification and data extraction
         Loan document classification workflow
         Loan data processing flow
      Media processing and analysis workflow
      E-commerce product recommendation
      Customer self-service automation with intelligent search
   Designing an MLOps architecture for AI services
      AWS account setup strategy for AI services and MLOps
      Code promotion across environments
      Monitoring operational metrics for AI services
   Hands-on lab – running ML tasks using AI services
      Summary
AI Risk Management
   Understanding AI risk scenarios
   The regulatory landscape around AI risk management
   Understanding AI risk management
      Governance oversight principles
      AI risk management framework
   Applying risk management across the AI lifecycle
      Business problem identification and definition
      Data acquisition and management
         Risk considerations
         Risk mitigations
      Experimentation and model development
         Risk considerations
         Risk mitigations
      AI system deployment and operations
         Risk considerations
         Risk mitigations
   Designing ML platforms with governance and risk management considerations
      Data and model documentation
      Lineage and reproducibility
      Observability and auditing
      Scalability and performance
      Data quality
   Summary
Bias, Explainability, Privacy, and Adversarial Attacks
   Understanding bias
   Understanding ML explainability
      LIME
      SHAP
   Understanding security and privacy-preserving ML
      Differential privacy
   Understanding adversarial attacks
      Evasion attacks
         PGD attacks
         HopSkipJump attacks
      Data poisoning attacks
         Clean-label backdoor attack
      Model extraction attack
      Attacks against generative AI models
      Defense against adversarial attacks
         Robustness-based methods
         Detector-based method
      Open-source tools for adversarial attacks and defenses
   Hands-on lab – detecting bias, explaining models, training privacy-preserving mode, and simulating adversarial attack
      Problem statement
      Detecting bias in the training dataset
      Explaining feature importance for a trained model
      Training privacy-preserving models
      Simulate a clean-label backdoor attack
   Summary
Charting the Course of Your ML Journey
   ML adoption stages
      Exploring AI/ML
      Disjointed AI/ML
      Integrated AI/ML
      Advanced AI/ML
   AI/ML maturity and assessment
      Technical maturity
      Business maturity
      Governance maturity
      Organization and talent maturity
      Maturity assessment and improvement process
   AI/ML operating models
      Centralized model
      Decentralized model
      Hub and spoke model
   Solving ML journey challenges
      Developing the AI vision and strategy
      Getting started with the first AI/ML initiative
      Solving scaling challenges with AI/ML adoption
         Solving ML use case scaling challenges
         Solving technology scaling challenges
         Solving governance scaling challenges
   Summary
Navigating the Generative AI Project Lifecycle
   The advancement and economic impact of generative AI
   What industries are doing with generative AI
      Financial services
      Healthcare and life sciences
      Media and entertainment
      Automotive and manufacturing
   The lifecycle of a generative AI project and the core technologies
      Business use case selection
      FM selection and evaluation
         Initial screening via manual assessment
         Automated model evaluation
         Human evaluation
         Assessing AI risks for FMs
         Other evaluation consideration
      Building FMs from scratch via pre-training
      Adaptation and customization
         Domain adaptation pre-training
         Fine-tuning
         Reinforcement learning from human feedback
         Prompt engineering
      Model management and deployment
   The limitations, risks, and challenges of adopting generative AI
   Summary
Designing Generative AI Platforms and Solutions
   Operational considerations for generative AI platforms and solutions
      New generative AI workflow and processes
      New technology components
      New roles
      Exploring generative AI platforms
         The prompt management component
         FM benchmark workbench
         Supervised fine-tuning and RLHF
         FM monitoring
   The retrieval-augmented generation pattern
      Open-source frameworks for RAG
         LangChain
         LlamaIndex
      Evaluating a RAG pipeline
      Advanced RAG patterns
      Designing a RAG architecture on AWS
   Choosing an LLM adaptation method
      Response quality
      Cost of the adaptation
      Implementation complexity
   Bringing it all together
   Considerations for deploying generative AI applications in production
      Model readiness
      Decision-making workflow
      Responsible AI assessment
      Guardrails in production environments
      External knowledge change management
   Practical generative AI business solutions
      Generative AI-powered semantic search engine
      Financial data analysis and research workflow
      Clinical trial recruiting workflow
      Media entertainment content creation workflow
      Car design workflow
      Contact center customer service operation
   Are we close to having artificial general intelligence?
      The symbolic approach
      The connectionist/neural network approach
      The neural-symbolic approach
   Summary
Other Books You May Enjoy
Index