Catalog of AWS machine learning, AI, generative AI, and big data/analytics services, plus supporting context: ML frameworks, hardware (Intel, GPUs, CUDA), and hands-on looks at QuickSight, Athena, and Glue.

AI vs ML vs Deep Learning

  • AI (Artificial Intelligence): machines that perform jobs that mimic human behavior 10:31:10
  • ML (Machine Learning): machines that get better at a task without explicit programming 10:31:10
  • DL (Deep Learning): machines using artificial neural networks inspired by the human brain to solve complex problems 10:31:10
  • Often shown as an "onion": AI ⊃ ML ⊃ DL (deep learning is ML using neural networks) 10:31:37

Amazon SageMaker (flagship ML service)

  • Fully managed service to build, train, and deploy ML models at scale; supports open-source frameworks like Apache MXNet (AWS-backed), TensorFlow, PyTorch, Hugging Face 10:31:37
  • SageMaker Ground Truth: data labeling service — humans label datasets used to train ML models 10:31:37
  • Amazon Augmented AI (A2I): human intervention review — when a SageMaker prediction has low confidence, it's queued for human review 10:32:29

AI/ML services catalog

  • Amazon CodeGuru: ML code analysis — performs code reviews, suggests code quality improvements, visual code profiles for performance 10:32:56
  • Amazon Lex: conversational interface service — build voice and text chatbots 10:32:56
  • Amazon Personalize: real-time recommendation service — same tech behind Amazon.com product recommendations 10:33:26
  • Amazon Polly: text-to-speech — upload text, get an audio file with a synthesized voice 10:33:26
  • Amazon Rekognition: image and video recognition — detect and label objects, people, celebrities 10:33:26
  • Amazon Transcribe: speech-to-text — upload audio, get text 10:33:57
  • Amazon Textract: OCR — extracts text from scanned documents/paper forms 10:33:57
  • Amazon Translate: neural machine translation using deep learning for more accurate, natural translations 10:33:57
  • Amazon Comprehend: NLP service — finds relationships in text (customer emails, support tickets, social media) to produce insights and predictions 10:34:28
  • Amazon Forecast: time-series forecasting — predicts business outcomes like product demand, resource needs, financial performance 10:34:28
  • AWS Deep Learning AMIs: EC2 images pre-installed with frameworks (TensorFlow, PyTorch, MXNet, Chainer, Gluon, Horovod, Keras) 10:34:58
  • AWS Deep Learning Containers: Docker images pre-installed with the same frameworks 10:34:58
  • AWS DeepComposer (ML musical keyboard), DeepLens (deep learning video camera), DeepRacer (ML-powered toy race car, raced at re:Invent) — all learning tools 10:35:31
  • Amazon Elastic Inference: attach low-cost GPU-powered acceleration to EC2 instances — reduces deep learning inference cost by 75% 10:35:31
  • Amazon Fraud Detector: fully managed fraud detection — identifies online payment fraud and fake account creation 10:36:01
  • Amazon Kendra: enterprise ML search engine — natural-language answers instead of simple keyword matching 10:36:27

Generative AI services (newer)

  • Generative AI: AI capable of generating new content — text, images, music, other media (e.g., Midjourney: prompt in, image out) 10:53:47
  • Amazon Bedrock: cloud service using large language models to generate text and image responses — "If you know what ChatGPT is, you know what Bedrock is" 10:51:28
  • Amazon CodeWhisperer: AI code generator, equivalent to GitHub Copilot 10:51:28
  • Amazon DevOps Guru: ML analysis of operational data/metrics/events to detect operational abnormalities 10:51:50
  • Amazon Lookout (3 offerings: Equipment, Metrics, Vision): quality control / automated inspections and anomaly detection 10:52:19
  • Amazon Monitron: predicts unplanned equipment downtime using IoT sensors capturing vibration/sensor data 10:52:51
  • AWS Neuron: SDK to run deep learning workloads on AWS Inferentia and Trainium (ML acceleration hardware) 10:52:51
  • Exam tip: Bedrock and CodeWhisperer are the ones most likely to appear on future exams 10:53:24

Big data and analytics services

  • Big data: massive volumes of structured/unstructured data too large to move and process with traditional database and software techniques 10:36:27
  • Amazon Athena: serverless interactive query service — run SQL queries over CSV/JSON files in S3 (loaded into a temporary SQL table); based on Apache Presto 10:36:52
  • Amazon CloudSearch: fully managed full-text search — add search to your website 10:36:52
  • Amazon Elasticsearch (ES): managed Elasticsearch cluster — more robust than CloudSearch but requires more server/operational maintenance 10:37:21
  • Amazon EMR (Elastic MapReduce): data processing and analysis — like Redshift for reports but better suited to transforming unstructured data into structured data on the fly; uses open source (Spark, Hive, Pig) 10:37:21
  • Kinesis Data Streams: real-time streaming data — producers send data to a stream, multiple consumers read it (real-time analytics, click streams, IoT fleet ingestion) 10:37:56
  • Kinesis Firehose: serverless, simpler version of a data stream — pay on demand per data consumed, no servers to manage 10:38:14
  • Kinesis Data Analytics: run queries against data flowing through a real-time stream — reports on emerging data 10:38:14
  • Kinesis Video Streams: analyze/process real-time streaming video 10:38:40
  • MSK (Managed Streaming for Kafka): fully managed Apache Kafka — real-time streaming data pipelines; similar to Kinesis but more robust 10:38:40
  • Redshift: petabyte-size data warehouse for OLAP (online analytical processing) — keeps data "hot" (expensive) so complex queries over large data return fast; for quickly generating analytics/reports 10:39:11
  • Amazon QuickSight: business intelligence (BI) dashboard — AWS equivalent of Tableau or Power BI; little/no programming, connects to many databases 10:39:38
  • AWS Data Pipeline: automates the movement of data between compute, storage, and services 10:39:38
  • AWS Glue: ETL service — move data with transformations before the final destination; similar to DMS but more robust 10:40:00
  • AWS Lake Formation: data lake — centralized, curated, secured repository holding vast raw data in native format until needed 10:40:00
  • AWS Data Exchange: catalog of third-party datasets (free, subscribe, or purchase) — e.g., COVID-19 foot traffic, IMDb, historical weather 10:40:47

Amazon QuickSight (detail + demo)

  • BI dashboard that ingests data from various AWS storage/database services with minimal programming; speed comes from SPICE — Super-fast Parallel In-memory Calculation Engine (optional but recommended) 10:40:47
  • Caveat: some data stores can't be ingested directly — dump to S3 first, optionally transform with Glue 10:41:13
  • QuickSight ML Insights: anomaly detection, forecasting, natural-language narratives (auto-written business reports) 10:41:37
  • QuickSight Q: ask questions in natural language over all your data, answers in seconds 10:41:37
  • Demo: separate sign-up (Standard vs Enterprise editions), integrates with Athena/S3/RDS 10:42:00; you can upload CSV/XLS files directly without S3 and drag fields to build bar/pie charts 10:45:40; to stop being billed you must delete the QuickSight account from account settings 10:50:38

ML / DL frameworks

  • Apache MXNet: the framework AWS adopted and markets heavily; in practice not pleasant to use 10:55:02 — deep learning framework supporting many programming languages; scalable, flexible, portable; strong support in SageMaker and ML containers 10:58:12; two high-level interfaces: Gluon (imperative) and Module API (symbolic) 10:58:38
  • PyTorch: created by Facebook; optimized for deep learning on GPU and CPU; well supported by all major providers 10:55:29
  • TensorFlow: made by Google, which built optimized hardware for it — the TPU (Tensor Processing Unit) 10:56:00
  • Keras: high-level framework built on top of TensorFlow, created to match PyTorch's ease of use 10:56:28
  • Apache Spark: unified analytics engine for large-scale data processing, with ML offering Spark ML 10:56:28
  • Chainer: deep learning framework supporting CUDA; Hugging Face: not a framework — a way to access many models/datasets online; strong AWS partnership 10:56:57
  • Exam tip: MXNet itself won't appear on the exam, but knowing it gives good AWS/ML context 10:59:32

Hardware: Intel, instruction sets, GPUs, CUDA

  • Intel: one of the world's largest semiconductor makers; inventor of the x86 instruction set (Intel 8086 chip, 1978) 11:00:02
  • When launching EC2 you choose the architecture: x86 or ARM; ARM has fewer instructions, so generally better power efficiency → better performance/cost — run ARM when your software supports it 11:00:58, 11:01:24
  • Intel Xeon Scalable processors: high-performance CPUs for enterprise/server use, common in AWS instances, good for ML 11:02:19
  • Intel Habana Gaudi: processor specialized for AI training (direct competitor to Nvidia), with its own SDK SynapseAI, usable via SageMaker 11:02:52
  • GPUs: specialized to render high-resolution images/video concurrently; parallel operations make them ideal for ML and scientific computation 11:03:54 — CPUs average 4–16 cores; GPUs can have thousands (up to ~40,000 cores across 48 GPUs); best for repetitive highly parallel tasks: graphics rendering, crypto mining, deep learning 11:04:51
  • Nvidia CUDA (Compute Unified Device Architecture): parallel computing platform/API allowing CUDA-enabled GPUs to be used for general-purpose computing (GPGPU) 11:05:48
  • AWS instances with Nvidia GPUs: P3 (Tesla V100), G3 (Tesla M60), G4 (T4), P4 (Tesla A100) 11:06:27; all major deep learning frameworks are integrated with Nvidia's Deep Learning SDK (e.g., cuDNN library) 11:06:58

Amazon Athena (deep dive)

  • Interactive query service to analyze data directly from S3 — the bucket is your dataset; based on a fork of the open-source distributed query engine Apache Presto 11:08:20
  • Two modes: Athena SQL (SQL queries over S3, uses Trino — a Presto fork; accessed via console, JDBC/ODBC drivers, CLI or SDK) 11:08:53 and Apache Spark on Athena (interactive data analytics in Jupyter-compatible notebooks) 11:09:38
  • Serverless — pay only for what you use; integrates with CloudFront, CloudTrail, DataZone, ELB, EMR, Glue Data Catalog, IAM, QuickSight, S3 Inventory, Step Functions, Systems Manager Inventory, VPC 11:10:05
  • Exam tip: knowing Athena's service integrations (what connects to it, where it dumps data) has mattered on exams like the security certification 11:10:05

AWS Glue (deep dive)

  • Serverless data integration service to discover, prepare, move, and integrate data from multiple sources — much improved since launch and worth knowing "inside and out" 11:10:46
  • Use cases: analytics, ML, application development; connects to 70+ data sources; centralized data catalog; visually create/run/monitor ETLs into data lakes 11:11:17
  • Catalog data is immediately searchable/queryable via Athena; strong integrations with EMR and Redshift Spectrum; does data discovery, modern ETL/ELT, cleansing, transforming, centralized cataloging 11:12:03