Introduction to Arti cial Intelligence
IFT3335 Lecture 1: Introduction
Bang Liu, Jian-Yun Nie fi
2 Outline
๏ Organization of the course
๏ Introduction to Artificial Intelligence (AI)
๏ Risks and Benefits of AI
Organization of the course
4 Teachers
๏ Professors:
• Bang Liu: [Link]@[Link] (English)
• Jian-Yun Nie: nie@[Link] (French)
๏ Teaching assistant:
• Hongjian Gu: [Link]@[Link] (French)
5 Course materials
๏ Documents on Studium
๏ Slack workspace: [Link]
qlAp9rp1Gn9_0RtICChxMg
๏ Course webpage:
•[Link]
(in progress)
๏ Book recommended for reference:
• Artificial Intelligence: A Modern Approach, 4th US ed. by Stuart Russell and Peter Norvig
• [Link]
6 Programming
๏ Programming language
• Python
๏ Resources
•[Link]
•[Link] (code)
7 Class Time
๏ Courses (January 12th ~ April 13th)
•Wednesdays 12:30 pm ~ 13:30 pm
•Thursdays 9:30 am ~ 11:30 am
๏ Practical Work (January 19th ~ April 13th)
• Wednesdays 13:30 pm ~ 15:30 pm
๏ Exams
•Mid-term: 12:30 ~ 14:30, March 9th
•Final: 12:30 ~ 15:30, April 27th
8 Evaluation
๏ Online exams
•Mid-term exam: 20% (2 hours)
•Final exam: 20% (2 hours)
๏ Small Exercises (30%)
•individual
•6 exercises on concepts or application of concepts (5% each)
๏ Practical work (programming) (30%)
• in groups of 2-3 people
• 2 programming labs (15% each)
9 Plan (tentative)
Topic Exercise (quiz) Practical work
1 Introduction and notion of agent (Chapter 1, 2)
2 Search in the state space (Chapter 3) Ex 1 (5%)
3 Two player games (Chapter 5) Ex 2 (5%)
4 Logic and knowledge (Chapter 7~10) Ex 3 (5%) TP1: Sudoku game (15%)
5 Uncertainty and Bayesian Networks (Chapter 12, 13)
Reading week: February 28th
Mid-term exam (March 9th) (20%)
6 Markov Decision and Reinforcement Learning (Chapter 16, 17) Ex 4 (5%)
7 Machine Learning (Chapter 19, 20) Ex 5 (5%)
8 Deep Learning (Chapter 21, 22) Ex 6 (5%)
9 Natural Language Processing (Chapter 23, 24) TP2: learning / NLP (15%)
10 Revision
Final exam (April 27th) (20%)
Introduction to Arti cial Intelligence (AI)
fi
What is AI ?
12 Imagination about AI
Bicentennial Man (1999, Chris Columbus) I, Robot (2004, Alex Proyas) Star Wars (1977, George Lucas)
Blade Runner 2049 (2017, Denis Villeneuve) Doraemon (1970, Fujiko Fujio) Astro Boy (1952, Osamu Tezuka)
13 Imagination about AI
14 What is AI?
๏ “the scientific understanding of the mechanisms underlying thought and intelligent behavior
and their embodiment in machines” (Association for the Advancement of Artificial Intelligence
– AAAI [Link]
• Understanding of intelligence
• Implantation in the machines
๏ Embodiment cognition
•a theory that many aspects of
cognition are shaped by the body
•The dictionary defines “to
embody” as “making visible”.
Our movement and body make
visible all of who we are: our
mood, personality, history, family
, and culture.
15 Levels of AI
Creative Intelligence
4 Imagination, creation
Cognitive Intelligence
3 Language, knowledge, reasoning
Perceptual Intelligence
2 Hearing, vision, touch
Pre-trained LMs
Computational Intelligence
1 Remember and calculate
16 Intelligence = ?
๏ Perceptions
๏ Execution capacity
๏ Learning
๏ Knowledge
๏ Reasoning
๏ Capacity for analysis, abstraction / generalization
๏ Adaptability
๏ Communications
๏ Intelligence quotient vs. Emotional Quotient
๏ ...?
17 What is AI?
๏ Humanly: behave like human beings
๏ Rationally: act like a rational person
๏ Thinking: mechanism of functioning
๏ Act: result
Image from: [Link]
18 Acting humanity: Turing test
๏ Produce responses like humans
๏ Turing (1950) “Computing machinery and
intelligence”
๏ "Can the machine think" → "Can the machine
behave intelligently?" (Like a human)
๏ Operational test for intelligent behavior: the
imitation game
๏ Question by the interrogator
๏ Response by human or AI system
๏ Can the interrogator distinguish between
the human and the system?
Image from: [Link]
19 Turing test
๏ First objective test
๏ Some limitations of the test
•Communication is done in natural language (weak point for the system)
•Limit oneself to human intelligence (e.g. unable to do complex calculations)
•Easy to trap the system
๏ He states that in 2000, a machine could have a 30% chance of cheating on a person for 5
minutes.
๏ The article anticipated all the major arguments against AI for the next 50 years
๏ Suggested the components of AI: knowledge, reasoning, language comprehension, learning
20 Variations and alternatives to the Turing Test
๏ Variations:
•Reverse Turing Test -- where a human tries to convince a computer that it is not a computer.
An example of this is a CAPTCHA.
•Total Turing Test -- where the questioner can also test perceptual abilities as well as the
ability to manipulate objects.
•Minimum Intelligent Signal Test -- where only true/false and yes/no questions are given.
๏ Alternatives:
•The Marcus Test -- in which a program that can 'watch' a television show is tested by being
asked meaningful questions about the show's content.
•The Lovelace Test 2.0 -- which is a test made to detect AI through examining its ability to
create art.
•Winograd Schema Challenge -- which is a test that asks multiple-choice questions in a
specific format.
21 Think humanly: cognitive modeling
๏ Systems should solve problems the same way humans do.
๏ “Cognitive revolution” in 1960s: psychology based on information processing
The cognitive revolution was a period during the 1950s-1960s when cognitive psychology
replaced Behaviorism and Psychoanalysis as the main approach in psychological fields. Increasing
focus was placed on observable behaviors in conjunction with brain activity and structure.
๏ Need scientific theories for internal brain activities.
๏ Two approaches:
•predict and test the behaviors of human subjects (top-down)
•or direct identification of neurological data (bottom-up)
๏ These 2 approaches (roughly speaking, cognitive science and cognitive neuroscience) are now
distinct fields of AI
22 Cognitive Science v.s. Neuroscience
23 Toy example: how computer works?
What is it, dude?
24 Neuroscience
I can see ….
If remove the machine is slow
Crack the hardware, but still don’t know how it works……
25 Cognitive Science
I can write words with
And I can draw with
Know what softwares can do, but don’t know how they works…
26 Cognitive Science v.s. Neuroscience
Neuroscience is the scienti c study of the
nervous system, aims to to understand the
fundamental and emergent properties of
neurons, glia and neural circuits
The goal of cognitive science is to
understand the principles of intelligence with
the hope that this will lead to a better
comprehension of the mind and of learning
and to develop intelligent devices.
Images from:
[Link]
[Link]
fi
27 Think rationally: “laws of thought”
๏ Aristotle: What are the correct argument / thought processes?
๏ Laws of thought, the three fundamental laws of logic that are ascribed to Aristotle:
1. The law of contradiction: for all propositions p, it is impossible for both p and not p to
be true
2. The law of excluded middle: either p or ∼p must be true, there being no third or middle
true proposition between them
3. The principle of identity: if a propositional function F is true of an individual variable x,
then F is true of x, or: F(x) ⊃ F(x), in which ⊃ means “formally implies.” Another
formulation of the principle of identity asserts that a thing is identical with itself, or (∀x)
(x = x), in which ∀ means “for every”; or simply that x is x.
28 Think rationally: “laws of thought”
๏ Some Greek schools have developed different forms of logic: notation and derivation rules for
thoughts; may or may not precede the idea of mechanization.
๏ Direct line from mathematics and philosophy to modern AI
๏ Problems:
•Not all intelligent behavior is the result of logical deliberation.
•What is the process of thinking? What thoughts should I be having?
29 Act rationally: rational agent
๏ Rational behavior: Doing Right Things
•Right things: those that are supposed to maximize the achievement of objectives, given the
information available.
๏ Rationality only concerns what decisions are made
(not the thought process behind them)
๏ Goals are expressed in terms of the utility of outcomes
๏ Being rational means maximizing your expected utility
30 Rational agent
๏ An agent is an entity which perceives and acts.
๏ This course aims to develop rational agents
๏ In a very abstract way, an agent is a function which maps sequences of perception to actions:
๏ For any class of environment and task, we aim to develop agents (or classes of agents) that
produce the best performance
๏ But computational limitations mean that perfect rationality is not attainable.
•→ Develop the best program for given machine resources
History and State-of-the-art of AI
32 AI Prehistory
33
A brief
history of AI
Deep Learning
34
35 Biological neurons
๏ Dendrite: Receives signals from
other neurons
๏ Soma: Processes the information
๏ Axon: Transmits the output of
this neuron
๏ Synapse: Point of connection to
other neurons
Biological Neurons:
An Overly Simpli ed Illustration
fi
36 The first computational model of a neuron
The first computational model of a neuron was
McCulloch-Pitts Neuron
proposed by Warren MuCulloch (neuroscientist) and
Walter Pitts (logician) in 1943.
37 Ancestor of neural networks: Perceptron
The "Perceptron" proposed by Cornell University
professor Frank Rosenblatt in 1957 was the first to use
algorithms to accurately define neural networks, and the
first mathematical model with self-organization and self-
learning capabilities. It is the ancestor of neural network
models. Frank Rosenblatt
38 Perceptrons can’t do XOR
1969: Marvin Minsky and Seymour
Pappert published a book - "Perceptron",
which cited some limitations in
Rosenblatt’s technique. They found that
perceptrons can’t do the Exclusive OR
logical operation.
Exclusive OR) is a digital logic gate that
gives a true (1 or HIGH) output when the
number of true inputs is odd. So, if input
0 and 1, or 1 and 0, the output is 1. In
other cases, the outputs are 0. This book
triggered a long AI Winter
[Link]
39 Second Generation NNs: Back-propagation
In July 1986, Hinton and David Rumelhart co-published
a paper in the journal Nature, "Learning
Representations by Back-propagating errors", which
was the first to systematically and concisely expound the
application of back-propagating algorithms to neural
network models. Neural network research begins to
re-arise.
[Link]
[Link]?origin=ppub
40 Convolutional Neural Networks
In the late 1990s, Yann Lecun used a
technique called “Convolutional
Neural Networks (CNN)” for
developing a commercial software to
read handwritten numbers on bank
checks. This check recognition system
occupied nearly 20% of the US market
in the late 1990s.
LeNet-5, convolutional neural networks
[Link]
41 Deep Belief Networks
[Link]
In 2006, Geoffrey Hinton introduced Deep Belief
Networks, also introduced layer-wise pretraining
technique, opened current deep learning era.
42 High-dimensional word embeddings and attention
In 2000, Yoshua Bengio authored the
landmark paper, “A Neural Probabilistic
Language Model,” that introduced high-
dimension word embeddings as a
representation of word meaning. Bengio’s
insights had a huge and lasting impact on
natural language processing tasks. His group
also introduced a form of attention
mechanism which led to breakthroughs in
machine translation and form a key
component of sequential processing with
[Link]
deep learning.
uploads/2019/08/[Link]
43 Generative Adversarial Networks
Since 2010, Bengio’s papers on generative
deep learning, in particular the Generative
Adversarial Networks (GANs) developed
with Ian Goodfellow, have spawned a
revolution in computer vision and computer
graphics. In one fascinating application of
this work, computers can actually create
original images, reminiscent of the creativity
that is considered a hallmark of human
intelligence. [Link]
44 Generative Adversarial Networks
45 ImageNet competition
At the end of 2012, Geoff Hinton and
his PhD students Alex Krizhevsky and
Ilya Sutskever took the first place in the
ImageNet image classification
competition, and increased the accuracy
rate to 84.7%.
Relying on deep learning, they shocked
the machine learning community. Since
then, a large number of researchers have
begun to enter this field.
In 2013, Google Hires Brains that
Helped Supercharge Machine Learning
46 More milestones
On the Origin of Deep Learning
47 2018 Turing Award
[Link]
AlphaGo vs. Lee Sedol
4-1
March 2016
48
49 From AlphaGo to MuZero
MuZero: Mastering Go, chess, shogi and Atari without rules
50 AlphaFold 2
AlphaFold: a solution to a 50-year-old grand challenge in biology
51 The neural history of NLP
Deep learning has brought great impacts to NLP
[Link]
52 OpenAI GPT-3
GPT-3, an autoregressive language model
with 175 billion parameters,
can "generate news articles which human
evaluators have difficulty distinguishing
from articles written by humans"
53 Automated driving
Waymo 360° Experience: A Fully Autonomous Driving Journey
54 Robots
55 Healthcare
The global Medical
Robotics market was 3980
million US$ in 2018 and is
expected to 12200 million US$
by the end of 2025, growing at
a CAGR of 15.0% between
2019 and 2025
[Link]
56 Virtual human
Microsoft XiaoIce 微软⼩冰 Xiaomi Xiaoai ⼩⽶⼩爱
Bilibili 泠鸢yousa Baidu Xiaodu 百度⼩度
The Era of Arti cial Intelligence
Network Manufacturing
Finance
Health Biology
Retail
HCI IoT Pre-trained LMs
Assistant
Transport Smart City Customer Service
57
fi
Risks and Bene ts of AI
fi
59 Benefits of AI
๏ Increase work efficiency
๏ Work with high accuracy
๏ Reduce cost of training and operation
๏ Accelerate scientific discovery
๏ “First solve AI, then use AI to solve everything else.” —- Demis Hassabis, CEO of DeepMind
60 Risks of AI
๏ Lethal autonomous weapons
๏ Biased decision making
๏ Safety-critical applications: driving cars, managing
water supplies, healthcare…
๏ Superintelligent: the gorilla problem
•What the gorilla problem means is loss of control.
Gorillas “created us”; but now we humans have
taken control and the gorillas have lost control. We
are now creating AI and lose control over it. To
make working AI, we have to hand over some
control. But: how much control is that? Do we cede
so much control to the AI that we turn from captains
to passengers of our ship?
61 References
๏ History of AI:
•[Link]
๏ [Link]
Thanks! Q&A
62
