Machine Learning on AWS

1. What this chapter covers

The deck’s Machine Learning section is a tour: one slide per service, and a closing slide that reduces each of them to a single line. Read it as a mapping exercise — for every service, the kind of problem it answers.

The order below is the deck’s own: Rekognition, Transcribe, Polly, Translate, Lex and Connect, Comprehend (with Comprehend Medical), SageMaker AI, Kendra, Personalize and Textract.

2. Amazon Rekognition – images and video

Amazon Rekognition uses ML to find objects, people, text and scenes in images and videos. It also performs facial analysis and facial search, which supports user verification and people counting. You can build a database of your own “familiar faces”, or compare against celebrities.

The use cases listed on the slide:

• Labeling — tagging the content of images and video.
• Content Moderation.
• Text Detection.
• Face Detection and Analysis — gender, age range, emotions and similar attributes.
• Face Search and Verification.
• Celebrity Recognition.
• Pathing — following movement, for example in sports game analysis.

Content Moderation

Rekognition Content Moderation detects content that is inappropriate, unwanted or offensive, in both images and videos. It is used in social media, broadcast media, advertising and e-commerce to create a safer user experience, and it helps comply with regulations.

How it works in practice:

• You set a Minimum Confidence Threshold for the items that will be flagged.
• Rekognition returns a confidence level, compared against that threshold.
• Sensitive content can be flagged for manual review in Amazon Augmented AI (A2I) — an optional human step after the automated pass.

Exam tip

When a scenario mentions filtering offensive user-uploaded images with a tunable confidence level and an optional human review, the pair to recognize is Rekognition Content Moderation plus Amazon Augmented AI (A2I).

3. Amazon Transcribe – speech to text

Amazon Transcribe takes audio in and gives text back, automatically. The deep learning technique behind it is automatic speech recognition (ASR), and the deck’s claim for it is that the conversion is both quick and accurate.

• It can automatically remove Personally Identifiable Information (PII) using Redaction.
• It supports Automatic Language Identification for multi-lingual audio.

Use cases from the slide:

• Transcribe customer service calls.
• Automate closed captioning and subtitling.
• Producing metadata for media assets, so that an archive of them becomes searchable end to end.

Real-world example

A support center records thousands of calls a day. Transcribe turns each recording into text with PII redacted automatically, so the transcripts can be stored and searched without holding customer identifiers in plain text — and the same transcripts become the input for sentiment analysis later in the pipeline.

4. Amazon Polly – text to speech

Amazon Polly turns text into lifelike speech using deep learning, letting you build applications that talk. It is the mirror image of Transcribe, and the exam usually tests only the direction of the conversion.

Lexicons and SSML

Two customization mechanisms are worth knowing:

• Pronunciation lexicons customize how words are pronounced. Examples from the slide: stylized words such as St3ph4ne read as “Stephane”, and acronyms such as AWS read as “Amazon Web Services”. You upload the lexicons and then use them in the SynthesizeSpeech operation.
• SSML (Speech Synthesis Markup Language) — you can generate speech from plain text, or from documents marked up with SSML, which enables more customization: emphasizing specific words or phrases, using phonetic pronunciation, including breathing sounds or whispering, and using the Newscaster speaking style.

Exam tip

Keep the two apart: a lexicon fixes how a particular word or acronym is pronounced, while SSML controls the delivery of the passage — emphasis, pauses, breathing, speaking style.

5. Amazon Translate – language translation

Amazon Translate provides natural and accurate language translation. It allows you to localize content — such as websites and applications — for international users, and to translate large volumes of text efficiently. When a question is about turning content from one language into another, this is the whole answer.

6. Amazon Lex and Amazon Connect – bots and contact centers

Amazon Lex uses the same technology that powers Alexa:

• Automatic Speech Recognition (ASR) to convert speech to text.
• Natural Language Understanding to recognize the intent of text and of callers.
• It helps build chatbots and call center bots.

Amazon Connect:

• Receives calls, lets you create contact flows, and is a cloud-based virtual contact center.
• It can integrate with other CRM systems or with AWS.
• No upfront payments, and it is 80% cheaper than traditional contact center solutions.

The two are usually shown together. A customer places a phone call, which Connect receives and streams to Lex; Lex recognizes the intent (“schedule an appointment”) and invokes a Lambda function, which schedules the appointment in the CRM.

Real-world example

A clinic replaces its phone menu with this chain: Connect answers, Lex understands “I would like to move my Tuesday appointment” without a keypad menu, Lambda updates the booking system, and only the calls Lex cannot handle reach a human agent.

7. Amazon Comprehend – natural language processing

Amazon Comprehend is the Natural Language Processing (NLP) service. It is fully managed and serverless, and it applies machine learning to find insights and relationships in text.

What it reads out of a document covers several layers at once. It identifies which language the text is written in. It pulls out the entities that matter — key phrases, places, people, brands and events. It scores the text for sentiment, that is, how positive or negative it reads. It breaks the text down grammatically, through tokenization and parts of speech. And across a whole collection of text files it will group them by topic on its own.

The slide’s two sample use cases follow from those: reading customer interactions such as emails to work out what drives a positive or negative experience, and grouping articles under the topics Comprehend discovers in them.

Amazon Comprehend Medical

Amazon Comprehend Medical is the clinical variant. It reads unstructured clinical text — physician’s notes, discharge summaries, test results, case notes — and returns the useful information it finds in them.

• It applies NLP to detect Protected Health Information (PHI), exposed through the DetectPHI API.
• The deck names three ways of getting text to it. Documents at rest can live in Amazon S3. A live feed can be analyzed as it arrives via Kinesis Data Firehose. And spoken material — a patient telling their story — can be put through Amazon Transcribe first, so that what Comprehend Medical analyzes is the transcript.

8. Amazon SageMaker AI – building your own models

Amazon SageMaker AI is a fully managed service for developers and data scientists to build ML models. The problem it solves is that doing the whole ML process in one place — and provisioning the servers for it — is typically difficult.

The deck illustrates the simplified machine learning process with the example of predicting an exam score:

1. Historical data with labels — number of years of experience in IT, number of years of experience with AWS, time spent on the course, and the score achieved (670, 890, 934).
2. Build and then train and tune an ML model from that labeled data.
3. Apply the model to new data to produce a prediction — “PASS WITH 906”.

Exam tip

SageMaker AI is the odd one out in this chapter. Every other service answers “give me a ready-made model behind an API”; SageMaker answers “I need to build, train and tune my own model”, which is exactly what the deck’s exam-score example walks through.

9. Amazon Kendra – ML-powered document search

Amazon Kendra is a fully managed document search service powered by Machine Learning.

• It extracts answers from within a document — text, PDF, HTML, PowerPoint, MS Word, FAQs and more.
• It offers natural language search capabilities: a user asks “Where is the IT support desk?” and gets the answer “1st floor” rather than a list of documents.
• It learns from user interactions and feedback to promote preferred results — Incremental Learning.
• You can manually fine-tune search results, adjusting the importance of data, freshness, custom criteria and so on.

It indexes from many data sources — Amazon S3, Amazon RDS, Google Drive, MS SharePoint, MS OneDrive, plus third-party, APN and custom sources — into a knowledge index powered by ML.

10. Amazon Personalize – real-time personalized recommendations

Amazon Personalize is a fully managed ML service to build applications with real-time personalized recommendations.

• Examples: personalized product recommendations and re-ranking, and customized direct marketing. If a user bought gardening tools, it recommends what to buy next.
• It is the same technology used by Amazon.com.
• It integrates into existing websites, applications, SMS and email marketing systems.
• You can implement it in days, not months — you do not need to build, train and deploy ML solutions yourself.
• Use cases: retail stores, media and entertainment.

In the architecture, Personalize reads data from Amazon S3 and also accepts real-time data integration through the Amazon Personalize API, then exposes a customized personalized API consumed by websites and apps, mobile apps, SMS and emails.

11. Amazon Textract – extracting data from scanned documents

Amazon Textract automatically extracts text, handwriting and data from any scanned document using AI and ML.

• It extracts data from forms and tables, returning structured output rather than a flat block of text — the slide shows fields such as Document ID, Name, SEX and DOB pulled out of an ID card as JSON.
• It can read and process any type of document — PDFs, images and so on.

Use cases:

• Financial services — invoices, financial reports.
• Healthcare — medical records, insurance claims.
• Public sector — tax forms, ID documents, passports.

Exam tip

Textract and Rekognition Text Detection both “read text”, which is exactly why they are paired as options. Rekognition finds text present in a photo or video scene; Textract extracts structured data from documents and forms, including handwriting and tables.

Quick recap

Service	What it does
Rekognition	Sees what is in an image or video — labels the content, detects and analyzes faces, picks out celebrities — plus content moderation with a confidence threshold and optional human review in A2I, text detection and pathing
Transcribe	Turns recorded audio into text, subtitles being the deck’s example, using ASR; PII Redaction and Automatic Language Identification
Polly	The other direction — text rendered as speech; pronunciation lexicons used in SynthesizeSpeech, and SSML for emphasis, phonetics, breathing and the Newscaster style
Translate	Moves text from one language into another — localizing websites and applications, and translating large volumes of text
Lex	The building block for conversational bots and chatbots; ASR plus Natural Language Understanding to recognize intent
Connect	A contact center that runs in the cloud; receives calls, contact flows, integrates with CRM, 80% cheaper than traditional solutions with no upfront payments
Comprehend	Reads meaning out of text (NLP) — language, key phrases, entities, sentiment, tokenization and parts of speech, topic grouping; Comprehend Medical detects PHI with DetectPHI
SageMaker AI	The service for developers and data scientists who need a model of their own — build it, train and tune it, then apply it to new data
Kendra	A document search engine with ML behind it, natural language answers, Incremental Learning from feedback, manual result tuning
Personalize	Recommendations tailored to each user and served in real time, the technology behind Amazon.com, implemented in days, integrated into sites, apps, SMS and email
Textract	Pulls text and data back out of documents — scanned forms, tables and handwriting, for financial services, healthcare and the public sector