Introduction to

Artificial
Intelligence

CS158-1
Prof. Raymond Sedilla
Artificial Intelligence &
Machine Learning
■ Define “Artificial Intelligence (AI),
“Machine Learning” (ML), and “Deep
Learning” (DL).
■ Explain how DL helps solve classical ML
limitations.
■ Explain key historical developments, and
the “hype – AI winter cycle”
■ Differentiate modern AI from prior AI.
AI Breakthroughs
Image Classification Machine Translation

“I love you” “Mahal kita”

As of 2016,
“Dog” “Cat”
Have achieved near – human performance
As of 2015,
Using the latest AI techniques
computers can be trained to perform
better on this task than humans
 ”About 100 years ago, electricity transformed
every major industry. AI has advanced to the
point where it has the power to transform…
AI is the every major sector in coming years.” – Andrew
Ng, Stanford University
New
Electricity *Projected Revenue from AI ($B) (Source: IDC)
is 95 Billion dollars in 2023
Relationship between AI, Machine and Deep
Learning

Artificial Intelligence

Machine Learning

DEEP LEARNING
Artificial
Intelligence “A BRANCH OF COMPUTER SCIENCE
DEALING WITH THE SIMULATION OF
INTELLIGENT BEHAVIOR IN
“COLLOQUIALLY, THE TERM ‘ARTIFICIAL
INTELLIGENCE’ IS APPLIED WHEN A
MACHINE MIMICS ‘COGNITIVE’
COMPUTERS.” – MERRIAM-WEBSTER FUNCTIONS THAT HUMANS ASSOCIATE
WITH OTHER HUMAN MINDS, SUCH AS
‘LEARNING’ AND PROBLEM SOLVING’.” -
WIKIPEDIA
Thinking Humanly Thinking Rationally

“The exciting new effort to make computers think ... machines “The study of mental faculties through the use of computational
with minds, in the full and literal sense.” (Haugeland, 1985) models.” (Charniak and McDermott, 1985)
“The study of mental faculties through the use of computational
models.” (Charniak and McDermott, 1985) “The study of the computations that make it possible to
perceive, reason, and act.” (Winston, 1992)
“[The automation of] activities that we associate with human
thinking, activities such as decision-making, problem solving,
learning ...” (Bellman, 1978)

Acting Humanly Acting Rationally

“The art of creating machines that perform functions that “Computational Intelligence is the study of the design of
require intelligence when performed by people.” (Kurzweil, intelligent agents.” (Poole et al., 1998)
1990) “Computational Intelligence is the study of the design of
intelligent agents.” (Poole et al., 1998) “AI . . . is concerned with intelligent behavior in artifacts.”
(Nilsson, 1998
“The study of how to make computers do things at which, at
the moment, people are better.” (Rich and Knight, 1991)

Figure 1.1 Some definitions of artificial intelligence, organized into four categories.

Artificial intelligence Modern Approach third edition (2016) - Russell and Norvig
Disciplines that gave ideas, viewpoint and strategies
to AI
1. Philosophy
2. Mathematics
3. Economics
4. Neuroscience
5. Psychology
6. Information Technology
7. Control theory and cybernetics
8. Linguistics
 The study and construction of
Machine programs that are not explicitly
programmed, but learn
Learning patterns as they are exposed
to more data over time
Machine Learning
These programs learn from repeatedly seeing the data, rather than being explicitly
programmed by humans.
Machine Learning
These programs learn from repeatedly seeing the data, rather than being explicitly
programmed by humans.

Emails are labeled

as spam vs not
Machine Learning
These programs learn from repeatedly seeing the data, rather than being explicitly
programmed by humans.

Machine
Learning
Program
Emails are labeled
as spam vs not the more emails
the program sees..
Machine Learning
These programs learn from repeatedly seeing the data, rather than being explicitly
programmed by humans.

Machine SPAM
Learning
Program NOT
Emails are labeled
as spam vs not the more emails SPAM
the program sees..
the better it gets
at classification
Machine Learning Terminology
In this example, we learn to classify flower species from a set of measurement
features.
Sepal length Sepal width Petal length Petal width species
6.7 3.0 5.2 2.3 virginica
6.4 2.8 5.6 2.1 virginica
4.6 3.4 1.4 0.3 setosa
6.9 3.1 4.9 1.5 versicolor
4.4 2.9 1.4 0.2 setosa
4.8 3.0 1.4 0.1 setosa
Machine Learning Terminology
In this example, we learn to classify flower species from a set of measurement
features.
Sepal length Sepal width Petal length Petal width species
Features
(attributes of the data) 6.7 3.0 5.2 2.3 virginica
6.4 2.8 5.6 2.1 virginica
4.6 3.4 1.4 0.3 setosa
6.9 3.1 4.9 1.5 versicolor
4.4 2.9 1.4 0.2 setosa
4.8 3.0 1.4 0.1 setosa
Machine Learning Terminology
In this example, we learn to classify flower species from a set of measurement
features.
Sepal length Sepal width Petal length Petal width species
Features
(attributes of the data) 6.7 3.0 5.2 2.3 virginica
6.4 2.8 5.6 2.1 virginica
Target
(column to be predicted) 4.6 3.4 1.4 0.3 setosa
6.9 3.1 4.9 1.5 versicolor
4.4 2.9 1.4 0.2 setosa
4.8 3.0 1.4 0.1 setosa
Two Main Types of Machine Learning
Dataset Goal Example

Supervised Has a Target Column Make Predictions

Fraud
Learning Detection

Unsupervised Does not have a Target Find Structure Customer

Learning Column in the Data Segmentation
Machine Learning
Limitations

• Suppose you wanted to

determine if an image is of a
car or a dog.
• What features would you use?
• This is where Deep Learning
can come in.
Deep learning

• Machine learning that involves

using very complicated models
called “Deep neural networks”
• Models determine best
representation of original data;
in classical machine learning,
humans must do this
Deep learning Example

Classic Machine
Feature
Machine Learning Raymond
Detection Classifier
Learning
Algorithm
Neural Network

Deep Learning
(Steps 1 and 2 are
combined into 1 step)
History of AI
AI has experienced several hype cycles ,where it has oscillated between periods of
excitement and disappointment

Early
Algorithms

“AI Winter” late

1960’s – 1970s
History of AI
AI has experienced several hype cycles ,where it has oscillated between periods of
excitement and disappointment

Early Expert Neural

Algorithms Systems Networks

“AI Winter” late “AI Winter” late

1960’s – 1970s 1980’s – 1990s
History of AI
AI has experienced several hype cycles ,where it has oscillated between periods of
excitement and disappointment

Early Expert Neural Machine Deep

Algorithms Systems Networks Learning Learning

“AI Winter” late “AI Winter” late

1960’s – 1970s 1980’s – 1990s
 Frank Rosenblatt
Alan Turing invented the
developed Turing perceptron
test, to test a algorithm. This was
machine’s ability to the precursor to
exhibit intelligent modern neural
behavior. networks.

1950 1957

1950s:
Early AI
1956 1959

Artificial Intelligence Arthur Samuel

was accepted as a published an
field at the Darmouth algorithm for a
Conference. checkers program
using machine
learning.
The First “AI
Winter”
1966: ALPAC committee evaluated AI techniques
for machine translation and determined there
was little yield from the investment
1969: Marvin Minsky published a book on the
limitations of the Perceptron algorithm which
slowed research in neural networks.
1973: The Lighthill report highlights AI’s failure to
live up to promises.

The two reports led to cuts in government

funding for AI research, leading to the first “AI
Winter”
 Expert Systems – systems with programmed rules
designed to mimic human experts.

Ran on mainframe computers with specialized

programming languages (e.g. LISP)
1980s:
AI Boom Were the first widely – used AI technology, with
two – thirds of “Fortune 500” companies using
them at their peak.

1986: The “Backpropagation” algorithm is able to

train multi – layer perceptron's, leading to new
successes and interest in neural network research
 Expert systems progress on solving business
problems slowed.

Another AI Expert systems began to melded into software suites

of general business applications (e.g. SAP, Oracle)
Winter that could run on PCs instead of mainframes

(1980s – Neural networks didn’t scale to large problems.

1990s)
Interest in AI in business declined.
 AI solutions had successes in speech
recognition, medical diagnosis, robotics,
and many other areas.

AI algorithms were integrated into larger

1990s – systems and became useful throughout
industry.
2000s:
Machine The Deep Blue chess system beat world
Learning chess champion Garry Kasparov.

Google’s search engine launched using

artificial intelligence technology.
2006: Rise of Deep
Learning
2006: Geoffrey Hinton publishes a paper on
unsupervised pre – training that allowed
deeper neural networks to be trained.
- Neural networks are rebranded as deep
learning.
2009: The ImageNet database of human
–tagged images is presented at the CVPR
conference.
2010: Algorithms compete on several visual
recognition tasks at the first ImageNet
competition.
Deep Learning
Breakthroughs (2012 - ?)

2012 : Deep learning beats previous benchmark on

the ImageNet competition.

2013: Deep learning is used to understand

“conceptual meaning” of words.

2014: Similar breakthroughs appeared in language

translation.

This have led to advancements in Web Search,

Document Search, Document Summarization, and
Machine Translation
Participants of Dartmouth Workshop reunited in
2006, after 50 years of the Dartmouth Workshop.
 Smart Finance
Industry
solutions Smart Healthcare

Security

Traffic

Games
Technologies
•Image recognition •Speech synthesis
•Image processing •Machine translation
•Video recognition •Sentiment analysis
•Speech recognition •Knowledge graph
•Semantic
understanding
 Computer vision

Speech processing

Technology Natural language processing

direction
Planning and decision-making system

Big data / Statistical analysis

 Machine Learning

Deep Learning

Algorithms Functional analysis

Computational mathematics

Probability theory, Fourier and wavelet

transform
AI application
scenarios

”Driverless Car”
AI application
scenarios

“Smart Home”
AI application
scenarios

“Virtual Reality”
AI application
scenarios

“Intelligent robot”
AI application
scenarios

“Smart
investment
adviser”
AI application
scenarios

“Smart
healthcare”
 • Speech signal processing
Speech signal

automatically and accurately

transcribes human speeches. A
complete speech signal processing
processing

system consists of signal processing,

speech recognition, semantic
recognition and dialogue
management, and speech synthesis.
 Processing Signal processing: human speech detection, echo
cancellation, wake-up-word recognition,
microphone array processing, speech enhancement,
etc.

Feature Speech recognition: feature extraction, model

adaptation, acoustic model, language model,
Speech dynamic decoding, etc.

signal Scope Semantic recognition and dialogue management:

processing scope of NLP

Text Speech synthesis: text analysis, linguistics analysis,

speech length estimation, vocal parameter
prediction, etc.
 Application: medical dictation,
speech dictation, voice
operated computer system,
Speech phone customer service, etc.
signal
processing Future: There is a long way to
go before machines can
communicate naturally with
people like human beings.
 Computer vision deals with how
computers can be made to
identify objects, scenes, and
Computer
activities from images, including
Vision image processing, recognition,
detection, analysis, and
understanding.
 Image processing: noise cancellation, deblurring,
super-resolution processing, filter processing, etc.

Image recognition: image pre-processing, image

Computer segmentation, feature extraction, and judgment and
matching. Image recognition can be used for
Vision classification, location, detection, and segmentation.

Image understanding: interaction between images

and texts. Image understanding can be used to
perform text-based image search, image description
generation, image Q&A, etc.
 IDENTIFYING SUSPECTS IN SECURITY
MEDICAL IMAGE ANALYSIS IS USED TO AND SURVEILLANCE FIELDS
FACILITATE DISEASE PREDICTION,
DIAGNOSIS, AND TREATMENT.

Computer
Vision
Applications

SHOPPING-GOERS CAN TAKE PHOTOS

OF PRODUCTS WITH SMARTPHONES
TO OBTAIN MORE INFORMATION.
 Computer vision is expected to
Future of enter an advanced stage of
independent understanding, and
analysis and decision making, truly
computer endow machines with the ability to
watch, and play a bigger role in
vision scenarios such as driverless cars and
smart home.
NLP
(Natural Language Processing)
NLP mainly involves knowledge
acquisition and expression, natural
language understanding, and
natural language generation. There
are also researches on knowledge
graph, dialogue management,
machine translation, etc.
NLP
(Natural Language Processing)
Knowledge graph: structured results obtained
after knowledge is organized based on
semantics
Dialog management: chatting, Q&A, and
task-driven dialogs

Machine translation: From traditional PBMT to

Google GNMT, the smoothness and accuracy
are greatly improved.
NLP
(Natural Language Processing)
Applications: search engine,
dialogue robot, machine
translation, college entrance
examination robot, intelligent
office secretary
 AI still in the initial
stage
AI is still in the initial stage and applicable to scenarios with known
environments, clear objectives, and predictable actions. Deep learning
deals with image recognition, speech recognition, and machine translation,
with AI having the same or even better recognition abilities compared with
human beings. These abilities are applied in many scenarios like healthcare
and public safety but are still weak in inference and cognition.

AI can come into use as long as it does better than human beings in a
certain aspect. We do not need to wait it to exceed human intelligence.
 AI is taking the landscape

AI might lead to a change in chip architectures, which will further reshape the industry
landscape. NVIDIA, Google, and Intel are competing for the dominant place in the future.
Might lead to a change in chip architectures, which will further reshape the industry
landscape. NVIDIA, Google, and Intel are competing for the dominant place in the future.
AI is the next major operating system. AI technology platforms will be a hot spot
competed by big players. AI applications or services of vertical industries might be
provided based on the technology platforms.
How AI Take Responsibilities?

At 22:00 on Sunday (March 19, 2018,

local time) in Tempe, Arizona, an
Uber's self-driving test car struck a
49-year-old woman, Elaine Herzberg,
who was crossing the street with a
bicycle. Elaine Herzberg was
transported to the hospital but later
died from her injuries. It was likely
the first time that a human
pedestrian has been killed by an
autonomous vehicle.
Copyright 2018 Huawei technologies Co. Ltd
 Efforts in self – driving cars
• In 2013, the U.S. National Highway Traffic Safety Administration
(NHTSA) issued the Federal Automated Vehicles Policy, which
stipulated the responsibilities in case of self-driving test
accidents.
• In August 2016, the United Nations Education, Scientific and
Cultural Organization (UNESCO) and World Commission on the
Ethics of Scientific Knowledge and Technology (COMEST)
explored the possibility of robots in the Preliminary Draft Report
of COMEST on Robotics Ethics. It suggested that people
participating in the robots' invention, authorization, and
distribution share the responsibility.
 Efforts in self – driving cars

• On May 12, 2017, Germany passed a revised action proposed

by the transport department. The system cannot completely
replace the driver for driving. The driver should stay there
and be able to take over the vehicle at any time. Although
self-driving is controlled by the computer, the ultimate
responsibility should primarily fall on the driver.
• Under the current legal framework, the robot is not liable for
any damage to the third party due to its behavior or
negligence.
 ▪ Legislation protection:
Legislation o Since the Swedish Data Protection Act,
the first personal data protection act,
Protection was issued in 1973, more than 110
and countries and regions have enacted
Technology dedicated laws to protect personal
information till December 2016.
Application o In 2012, Decision of the Standing
Committee of the National People's
Congress on Strengthening Information
Protection on Networks was passed. In
2016, the People's Republic of China
Network Security Law was adopted.
Legislation
Protection ▪Technical application:
and oData anonymization: It is the
Technology process of removing personally
identifiable information from
Application personal data, so that the people
whom the data describes remain
anonymous.
 ▪ In many cases, algorithm-based
decision-making is to use the past data to
predict the future. The algorithm model
and input data, which determine the
Is Algorithm prediction results, are two main sources
of algorithm discrimination. The following
Fair by are two examples:
o Google's image software once marked
default? a black person as a gorilla by mistake.
o On March 23, Microsoft's AI chatbot,
Tay, was taught to be an anti-Semite,
sexist, and racist after it was launched.
As a result, it was brought offline
immediately in less than a day.
Issues to be resolved

•Are the contents

created by AI protected
by copyright laws?
•Who will assign rights
to robots?
•What rights can be
assigned to robots?
Three generations of
robots
• Generation 1: Playback robot. It can repeat
actions taught by humans but is unaware of
the outside environment.
• Generation 2: Robot with feelings. It has
feelings similar to humans. For example, it can
judge a force through the sense of force,
touch, and hearing.
• Generation 3: Intelligent robot. It is the
highest pursuit of robot development. An
intelligent robot is expected to be able to do
what people ask it to. Now it remains to be a
concept.
Classification of
Intelligent Robots
▪ There is no unified definition of AI
research in the world. Currently,
intelligent machines are classified
into four types:
"Think like people": Weak AI, such
as Watson and AlphaGo
"Act like people": Weak AI, such as
Android, iRobot, and Atlas of
Boston Dynamics
"Think rationally": Strong AI,
which is yet unavailable due to the
bottlenecks in brain science
"Act rationally": Strong AI
 • In 1942, Isaac Asimov, a well-known
American science fiction author, proposed
the Three Laws of Robotics.
Law One – “A robot may not injure a human
Three Laws being or, through inaction, allow a human
being to come to harm.”
of Robotics Law Two – “A robot must obey orders given
to it by human beings except where such
orders would conflict with the First Law.”
Law Three – “A robot must protect its own
existence, as long as such protection does not
conflict with the First or Second Law.”
 • Man-machine relationship
blueprint:
There are worries that robots
might pose threats to human
Man-Machine beings. However, machines and
human beings can coexist by
Relationship and controlling AI.
AI Governance AI becomes the agent of human
consciousness. Human beings
extend themselves through AI.
The virtual reality will come true
in the future.
 • AI governance:
AI governance should be based
on technological and industrial
innovation.
Regulators are advised to give
Man-Machine more freedom to the market for
Relationship and innovation.
Do not set too many constraints
AI Governance on the grounds of security.
Strive to facilitate development
and innovation.
Encourage different entities to
participate in the AI governance.
 • Opportunities: From efficiency to
intelligence, AI will create a
Opportunities market larger than today's IT
market (US$2 trillion), which
and Challenges sparks a competition in the
information industry.
of AI:
• Challenges: In the entire industry
New Markets chain, people who master
intelligence will have greater say
and Value and gain more value. This is why
Distribution traditional enterprises, such as
GE, set up their own digital
departments.