AI L3: AI Around Us

Recommendation system

A subclass of information-filtering system that predicts which items a user is most likely to want and surfaces them ahead of the rest of the catalog

Use of Recommendation system

They filter huge catalogs of songs, videos, products, and other content into a short, personally relevant list

Content-based filtering

A recommendation approach that compares the attributes of items you engaged with and suggests similar items

Key phrase for Content-based filtering

"More of what you already liked."

What attributes can content-based filtering compare?

Genre, artist, keywords, and other item attributes

Example of content-based filtering

Watching a space documentary leads YouTube to suggest more space documentaries

Collaborative filtering

A recommendation approach that finds users with similar rating or behavior patterns and recommends things they liked that you haven't tried

Key phrase for Collaborative filtering

"People like you also liked..."

Example of collaborative filtering

Shopee or Lazada's "you may also like" product recommendations

What user data signals can recommendation systems use?

Clicks, watch time, likes, ratings, purchase history, and search terms

Examples of real-world recommendation systems

Netflix and YouTube's "Because you watched..." recommendations, Spotify's Discover Weekly and Daily Mixes, and Shopee/Lazada's "you may also like" suggestions

Virtual assistant

An AI-driven application that understands natural-language commands, typed or spoken, and carries out tasks or answers questions on the user's behalf

3 examples of Virtual Assistants

Siri, Google Assistant, and Alexa

4 stages of Virtual Assistant’s Pipeline

• Speech Recognition

• Intent Understanding

• Action/Response

• Speech Synthesis

What happens during Speech Recognition?

Audio is converted into text.

What happens during Intent Understanding?

NLP identifies the goal of the user's request.

What does NLP stand for?

Natural Language Processing

Natural Language Processing (NLP)

The field that lets computers work with human language

What happens during Action/Response?

The assistant executes the task

What happens during Speech Synthesis?

A spoken confirmation is generated

What tasks can virtual assistants perform?

Setting reminders and alarms, web and knowledge searches, controlling smart-home devices, and real-time translation

Order of the Virtual Assistant Pipeline

Speech Recognition → Intent Understanding → Action/Response → Speech Synthesis

Chatbot

A program that carries on a text or voice conversation with a user

2 Types of Chatbots

Rule-Based Chatbots and AI/NLP-Driven Chatbots

Rule-based Chatbot

A chatbot that follows a fixed decision tree of pre-written questions and answers

Advantage of Rule-based chatbots?

They are predictable

Limitation of Rule-based chatbots

They can break down when a question falls outside the script

Example of Rule-based chatbot

Bank hotline menu-style bots

AI/NLP-driven chatbot

A chatbot that uses NLP to interpret meaning and generate flexible responses

Advantage of AI/NLP-driven chatbots

They can respond to phrasing they haven't seen exactly before

Example of AI/NLP-driven chatbot

Modern customer-service assistants

What was ELIZA?

A chatbot built by Joseph Weizenbaum of MIT in 1966, widely regarded as the first chatbot

When was ELIZA created?

1966

Who built ELIZA?

Joseph Weizenbaum of MIT

What did ELIZA simulate?

A Rogerian psychotherapist

How did ELIZA generate replies?

It scanned user input for keywords and applied decomposition and reassembly rules

ELIZA

No real language understanding

What idea do modern AI chatbots build upon from ELIZA?

Recognizing patterns in input, but on a vastly larger scale

What are modern AI chatbots trained on?

Huge collections of text

What do modern AI chatbots use text to do?

Recognize patterns and generate contextually relevant replies

Facial Recognition

A system that maps and compares the geometry of a face to confirm or determine identity

4 Stages of Facial Recognition

• Image Capture

• Feature Extraction

• Comparison/Matching

• Classification/Verification

What happens during Image Capture?

A camera captures a photo or video frame of a face

What happens during Feature Extraction?

Facial landmarks such as eye spacing, jawline, and nose shape are mapped

What happens during Comparison/Matching?

Features become a numeric template and are compared

What happens during Classification/Verification?

The system outputs a match decision or identity label

Facial Verification

A 1:1 comparison that asks, "Is this the same face?"

Example of Verification

Unlocking a smartphone with Face ID

Facial Identification

A 1:N comparison that asks, "Whose face is this?"

Example of Identification

Auto-tagging friends in a social media photo or airport security checkpoints

Difference between Verification and Identification

Verification compares one face with one stored template, while identification compares one face against an entire database of many faces

What does 1:1 mean in facial recognition?

One face is compared with one stored template

What does 1:N mean in facial recognition?

One face is compared with many faces in a database

Five benefits of everyday AI applications

Convenience, personalization, accessibility, security, and efficiency

How does AI provide convenience?

Through hands-free tasks, instant answers, and one-tap purchases

How does AI provide personalization?

By tailoring content and products to individual taste

How does AI improve accessibility?

Voice control and captions can widen access for users with disabilities

How can AI improve security?

Biometric verification can be faster and harder to fake than a password

How does AI improve efficiency?

It can reduce search and decision time across huge catalogs

4 risks of everyday AI applications

Privacy and data collection, algorithmic bias, filter bubbles, and over-reliance

What is the risk of privacy and data collection?

Clicks, likes, and face scans can feed a profile that the user may not control.

Algorithmic bias

Differences in AI system performance or accuracy across demographic groups

Filter bubbles

Situations where recommendation systems can narrow, rather than widen, what a user sees over time

Over-reliance on AI

Depending too heavily on AI answers, which can reduce our own critical thinking and verification habits.

Gender Shades study

A 2018 study by Buolamwini and Gebru that evaluated three commercial gender-classification systems

What did the Gender Shades study find?

Darker-skinned women were misclassified at much higher rates than lighter-skinned men

What does the Gender Shades study demonstrate?

Facial recognition accuracy is not evenly distributed across users and can demonstrate algorithmic bias

What should you check before granting an app access to personal features?

App permissions, especially camera, microphone, and location access

What should you ask about a recommendation?

Why you are seeing it, not just what it is

How should you protect personal data when using AI?

Avoid oversharing personal information with chatbots and virtual assistants

When should you verify AI-driven decisions?

Especially when they affect access or identity