Chapter 1 - Introduction to Artificial Intelligence (AI)

What is Artificial Intelligence (AI)?

AI is the technology that enables computers or machines to simulate human-like

intelligence, such as seeing, hearing, speaking, interacting, thinking, predicting,
making decisions, etc).

Some common abilities of AI technologies

AI enables computers or machines to:

● Communicate using human language

● Form concepts
● Self learn and self improve
● Simulate human logic to solve problems

Examples of AI technologies

Self-driving cars: Computer vision

Facial recognition: Computer vision
technologies are used in self-driving
techniques are used to detect (see)
cars to see the road and make
faces in images.
corresponding decisions.

1
Chatbots: Natural language processing Song recommendation systems: Deep
technologies are used to understand learning is used to recommend songs
user inputs and generate responses based on user history (decision making).
(communicate using human language).

Voice-controlled robots: Speech recognition is also applied to process voice

statements (hearing) and Natural Language Processing is applied to understand
the commands.

Examples of Applications/Technologies that do not use AI

Automated Teller
3D printer Web browsing
Machine (ATM)

2
Chapter 2 - Fundamentals of AI

What is an AI model?

AI models are based on mathematical algorithms that use a large amount of input
data for training and once trained, simulate the decisions made by experts.

For example, given a set of input data such as images of cats and dogs, the trained
AI model is able to classify the images as “cats” and “dogs”.

Model Training

In order to train an AI model for a specific task, we need a large amount of labeled
data. During training, the AI model learns to extract key features from the data.

3
Once trained, the model can make predictions on new data.

Impact of Training Data on Model Performance

Training data has a critical influence on an AI system. The quantity and quality of
the training data affects the performance of the AI model.

● Quantity
○ More training data leads to higher accuracy
○ Insufficient data, including insufficient data coverage, leads to
incorrect results, e.g., a lack of data for Cantonese in Text-to-Speech
● Quality
○ High quality data leads to a high performing model
○ Low quality data leads to a low performing model
■ Biased data – imbalance between different classes of training
data (the model will pay more attention to the majority class)
■ Noisy data – training data is incorrectly labeled

4
Perceptrons and Neural Networks

Artificial Neural Networks are modeled after the human brain. Similar to how
neurons are the basic units of the brain and the nervous system, an artificial neural
network is made up of neurons called perceptrons.

A perceptron is able to generate output based on the given input(s) and its
associated internal parameter(s).

For complex tasks, multiple perceptrons are connected in layers to form a neural
network, where the output of a previous layer is passed on as the input to the next
layer.

During training, the internal parameters of each perceptron in the neural network
are updated at the end of each epoch according to the learning rate.

● If the learning rate is set too high, the model may learn faster, however, it
may produce a less-than-optimal model.
● If the learning rate is set too low, training will take too long.

5
Chapter 3 - See

What is Computer Vision (CV)?

Computer Vision (CV) is a field of AI that enables computers and systems to derive
meaningful information from digital images, videos, and other visual inputs – and
take actions or make recommendations based on that information.

In other words, computer vision allows computers to see, observe, and understand
the world. Computer vision ‘sees’ the world through sensors and cameras, and
analyzes it with the use of algorithms to perform tasks such as object detection,
classification, tracking, etc.

Sense

Analyze

When adopting computer vision solutions, it is important to pay attention to privacy

and ethical issues that may arise.

6
Chapter 4 - Hear

What is Automatic Speech Recognition (ASR)?

Automatic speech recognition (ASR) is a technology that enables computers to

recognize and translate human spoken speech into text, i.e., speech to text.

ASR technology involves training and testing an AI model.

Once model training is completed, the model will interpret the input speech and
output the transcription.

The input for an ASR system, i.e., spoken audio, is captured using acoustic devices
such as microphones.

7
 ● While ASR technology is quite mature now, the accuracy of recognition and
transcription is still affected by various factors:

○ Speaker characteristics – speaking style, age, accent, vocabulary, phrasing,

pauses

○ Accent – regional trends cause the same word to sounds different

(e.g., American and Australian accent for ‘data’)

○ Environmental factors – background noises

○ Recording device – available quality

○ Low resource languages – lack of training data for a given language

8
Chapter 6 - Read

What is Natural Language Processing (NLP)?

Natural language processing (NLP) is a branch of AI that helps computers

understand, interpret, and manipulate human language, either in text or speech
form.

NLP Applications

NLP applications are widely used in our daily lives. The following are some
examples of NLP technologies:

Sentiment Analysis is used to recognize subtle nuances in emotions and opinions ‒

and determine how positive or negative they are.

Machine translation is the process of using artificial intelligence to automatically

translate text from one language to another without human involvement.

9
Chatbots and Virtual Assistants are used for automatic question answering,
designed to understand natural language and deliver an appropriate response
through natural language generation.

10
Chapter 7 - AI Reasoning

What are reasoning systems?

A reasoning system is a system that uses logical techniques, such as induction and
deduction, to generate conclusions from a knowledge dataset.

They have a wide range of applications, such as scheduling, rule processing,

problem solving, complex event processing, intrusion detection, robotics, computer
vision and natural language processing.

Levels of reasoning

Reasoning can be loosely grouped into 3 levels:

1. Skills-based reasoning: actions taken based on intuitive sensory-motor

behavior, without conscious control of straightforward tasks with mechanical
movements.

2. Rules-based reasoning: actions based on stored rules or procedures derived

from prior occasions or other’s experiences for well-defined and structured
tasks with a small degree of adaptation and planning.

3. Knowledge-based reasoning: actions derived from adapting previous

knowledge and planning, or analyzing the tasks to achieve a goal for complex
tasks in uncertain or unfamiliar environments.

11
Chapter 8 - Simulation

What is Simulation?

Simulation imitates real-world systems or processes by including real-world

components and real-world rules (the inner mechanisms of the real-world system).
Simulation is continuous in time, and the speed of simulation can be tuned faster,
slower, or paused. Examples of simulations include driving simulation, flight
simulation, and even games.

Simulation allows us to explore situations and do experiments that would

otherwise be difficult or impossible to do so that we can better understand the
situation. And thus, simulation can also help humans make better decisions.

Building a simulation environment

Step 1: Preparing the world model

To create a driving simulation, we first need to build and set parameters for objects
such as roads, traffic lights, driving rules, speed limits and pedestrian crossings.

Step 2: Analyzing and constructing the physics

After building the world model, we need to work on modeling the physics of some
objects in the simulation environment, such as car engines, transmissions, tires,
grips, steering and aerodynamics.

Step 3: Enable controls and interaction

The last step would be to allow users to control or interact with objects in the
simulation environment, such as steering wheels, throttle pedals, brake pedals,
gearboxes and dashboards.

12
We also need to set the responses of some objects in the world model, such as
collision simulations, traffic light interactions and speed limits.

How AI learns from simulation

Similar to how humans learn something new, AI can learn from two approaches –
supervised learning (SL) and reinforcement learning (RL).

Supervised learning is mainly about using data that is well labeled, i.e., data that is
already tagged with the correct answer. For example, in the images of cats and
dogs that we went through above, each image in the training dataset was already
labeled as “cat” or “dog” .

Reinforcement learning is based on the simple principle of trial-and-error, where

the AI model learns by itself by performing different actions and distinguishing
between good and bad actions.

13
Chapter 11 - AI and Ethics

AI Ethical Principles

1. Transparency – AI technologies must be understood and explained to users.

2. Justice and Fairness – AI must respect equality, inclusion, and diversity.

3. Beneficence – AI must do no harm.

4. Responsibility – AI must be accountable and honest.

5. Privacy – AI must protect and secure sensitive data.

14