1.

Intelligent machines,
Lecture 1 or what machines can do?
From intelligent machine to machine learning • Philosophers have been trying for over 2000 years to
understand and resolve two Big Questions of the Universe:
1. Intelligent machines, or what machines can do? How does a human mind work, and Can non-humans have
minds? These questions are still not well answered.
2. Turing test (Here is one answer https://www.newworldai.com/how-to-create-a-mind-the-secret-
of-human-thought-revealed/)
3. The history of AI: from the “Dark Ages” to deep
learning era • Intelligence is the ability to understand and learn things.
2 Intelligence is the ability to think and understand instead of
4. Machine Learning
doing things by instinct or automatically.
(Essential English Dictionary, Collins, London, 1990)

1 2

• In order to think, someone or something

must have a brain, or an organ that enables • Some people are smarter in some ways than others.
someone or something to learn and • Sometimes we make very intelligent decisions but sometimes
understand things, to solve problems and to we also make very silly mistakes.
make decisions. So we can define
intelligence as the ability to learn and • Some of us deal with complex mathematical and engineering
understand, to solve problems and to make problems but are moronic in philosophy and history.
decisions. • Some people are good at making money, while others are
• The goal of artificial intelligence (AI) as a better at spending it.
science is to make machines do things that • As humans, we all can learn and understand, to solve problems
would require intelligence if done by and to make decisions; however, our abilities are not equal
humans. Therefore, the answer to the and lie in different areas.
question Can Machines Think? was vitally
important to the discipline. • Therefore, we should expect that if machines can think, some
of them might be smarter than others in some ways.
• The answer is not a simple “Yes” or “ No ”.
3 4
 2. Turing Test Turing Imitation Game: Phase 1

• One of the most significant papers on machine intelligence, Turing did not provide definitions
“Computing Machinery and Intelligence”, was written by of machines and thinking, he just
the British mathematician Alan Turing over fifty years ago. avoided semantic arguments by
However, it still stands up well under the test of time, and the inventing a game, the Turing
Turing’s approach remains universal. Imitation Game.

• He asked: Is there thought without experience? Is there

mind without communication? Is there language without
living? Is there intelligence without life? All these questions, The imitation game originally included two phases. In the first phase, the interrogator,
a man and a woman are each placed in separate rooms. The interrogator’s objective is
as you can see, are just variations on the fundamental to work out who is the man and who is the woman by questioning them. The man
question of artificial intelligence, Can machines think? should attempt to deceive the interrogator that he is the woman, while the woman must
convince the interrogator that she is the woman.

5 6

Turing Imitation Game: Phase 2

The Turing test has two remarkable qualities that
make it universal.
In the second phase of the game,
the man is replaced by a computer
programmed to deceive the • By maintaining communication between the human and the
interrogator as the man did. It machine via terminals, the test gives us an objective standard
would even be programmed to view on intelligence.
make mistakes and provide fuzzy • The test itself is quite independent from the details of the
answers in the way a human experiment. It can be conducted as a two-phase game, or even
would. If the computer can fool as a single-phase game when the interrogator needs to choose
the interrogator as often as the between the human and the machine from the beginning of the
man did, we may say this test.
computer has passed the
intelligent behavior test.

7 8
 3. The history of artificial intelligence
• Turing believed that by the end of the 20th century it would be
possible to program a digital computer to play the imitation The birth of artificial intelligence (1943 – 1956)
game. Although modern computers still cannot pass the Turing
test, it provides a basis for the verification and validation of • The first work recognized in the field of AI
knowledge-based systems. was presented by Warren McCulloch and
• A program thought intelligent in some narrow area of Walter Pitts in 1943. They proposed a
expertise is evaluated by comparing its performance with model of an artificial neural network and
the performance of a human expert. demonstrated that simple network
• To build an intelligent computer system, we have to capture, structures could learn.
organize and use human expert knowledge in some narrow • McCulloch, the second “founding father”
area of expertise. of AI after Alan Turing, had created the
corner stone of neural computing and
artificial neural networks (ANN).
9 10

• The third founder of AI was John von Neumann, the brilliant • Another of the first-generation researchers was Claude
Hungarian-born mathematician. In 1930, he joined the Princeton Shannon. He graduated from MIT and joined Bell Telephone
University, lecturing in mathematical physics. He was an adviser Laboratories in 1941. Shannon shared Alan Turing’s ideas on
for the Electronic Numerical Integrator and Calculator project at the possibility of machine intelligence. In 1950, he published
the University of Pennsylvania and helped to design the a paper on chess-playing machines, which pointed out that a
Electronic Discrete Variable Calculator. He was influenced by typical chess game involved about 10120 possible moves
McCulloch and Pitts’s neural network model. When Marvin (Shannon, 1950). Even if the new von Neumann-type
Minsky and Dean Edmonds, two graduate students in the computer could examine one move per microsecond, it would
Princeton mathematics department, built the first neural network take 3 ´ 10106 years to make its first move. Thus Shannon
computer in 1951, von Neumann encouraged and supported demonstrated the need to use heuristics in the search for the
them. solution.

11 12
• In 1956, John McCarthy, Martin Minsky and Claude The rise of artificial intelligence, or the era of
Shannon organized a summer workshop at Dartmouth College. great expectations (1956 – late 1960s)
They brought together researchers interested in the study of
machine intelligence, artificial neural nets and automata theory.
Although there were just ten researchers, this workshop gave • The early works on neural computing
birth to a new science called artificial intelligence. and artificial neural networks started
by McCulloch and Pitts was
continued. Learning methods were
improved, and Frank Rosenblatt
proved the perceptron convergence
theorem, demonstrating that his
learning algorithm could adjust the
connection strengths of a perceptron.

13 14

• One of the most ambitious projects of the era of great • However, GPS failed to solve complex problems. The program
expectations was the General Problem Solver (GPS). was based on formal logic and could generate an infinite
Allen Newell and Herbert Simon from the Carnegie Mellon number of possible operators. The amount of computer time
University developed a general-purpose program to simulate and memory that GPS required to solve real-world problems
human-solving methods. led to the project being abandoned.
• Newell and Simon postulated that a problem to be solved • In the sixties, AI researchers attempted to simulate the thinking
could be defined in terms of states. They used the mean-end process by inventing general methods for solving broad classes
analysis to determine a difference between the current and of problems. They used the general-purpose search mechanism
desirable or goal state of the problem, and to choose and to find a solution to the problem. Such approaches, now
apply operators to reach the goal state. The set of operators referred to as weak methods, applied weak information about
determined the solution plan. the problem domain.

15 16
 Unfulfilled promises, or the impact of reality
(late 1960s – early 1970s)

The main difficulties for AI in the late 1960s were:

• By 1970, the euphoria about AI was gone, and most government
funding for AI projects was cancelled. AI was still a relatively • Because AI researchers were developing general methods for
new field, academic in nature, with few practical applications broad classes of problems, early programs contained little or even
apart from playing games. So, to the outsider, the achieved no knowledge about a problem domain. To solve problems,
results would be seen as toys, as no AI system at that time could programs applied a search strategy by trying out different
combinations of small steps, until the right one was found. This
manage real-world problems.
approach was quite feasible for simple toy problems, so it
seemed reasonable that, if the programs could be “scaled up” to
solve large problems, they would finally succeed.

• 17 18

• In 1971, the British government also suspended support for AI

• Many of the problems that AI attempted to solve were too research. Sir James Lighthill had been commissioned by the
broad and too difficult. A typical task for early AI was Science Research Council of Great Britain to review the
machine translation. For example, the National Research current state of AI. He did not find any major or even
Council, USA, funded the translation of Russian scientific significant results from AI research, and therefore saw no need
papers after the launch of the first artificial satellite (Sputnik) to have a separate science called “artificial intelligence”.
in 1957. Initially, the project team tried simply replacing
Russian words with English, using an electronic dictionary. • https://www.youtube.com/watch?v=03p2CADwGF8
However, it was soon found that translation requires a general
understanding of the subject to choose the correct words. This
task was too difficult. In 1966, all translation projects funded
by the US government were cancelled.

19 20
 The technology of expert systems, or the key to
success (early 1970s – mid-1980s)
DENDRAL
• Probably the most important development in the seventies was
• DENDRAL was developed at Stanford University to determine
the realization that the domain for intelligent machines had to be
the molecular structure of Martian soil, based on the mass
sufficiently restricted. Previously, AI researchers had believed
spectral data provided by a mass spectrometer. The project was
that clever search algorithms and reasoning techniques could be
supported by NASA. Edward Feigenbaum, Bruce Buchanan (a
invented to emulate general, human-like, problem-solving
computer scientist) and Joshua Lederberg (a Nobel prize winner
methods. A general-purpose search mechanism could rely on
in genetics) formed a team.
elementary reasoning steps to find complete solutions and could
use weak knowledge about domain. • There was no scientific algorithm for mapping the mass spectrum
into its molecular structure. Feigenbaum’s job was to incorporate
When weak methods failed, researchers finally realized that the the expertise of Lederberg into a computer program to make it
only way to deliver practical results was to solve typical cases in perform at a human expert level. Such programs were later
narrow areas of expertise, making large reasoning steps. called expert systems.
21 22

• DENDRAL marked a major “paradigm shift” in AI: a shift from

general-purpose, knowledge-sparse weak methods to domain- • A 1986 survey reported a remarkable number of successful
specific, knowledge-intensive techniques. expert system applications in different areas: chemistry,
• The aim of the project was to develop a computer program to electronics, engineering, geology, management, medicine,
attain the level of performance of an experienced human chemist. process control and military science (Waterman, 1986).
Using heuristics in the form of high-quality specific rules, rules- Although Waterman found nearly 200 expert systems, most
of-thumb, the DENDRAL team proved that computers could of the applications were in the field of medical diagnosis. Seven
equal an expert in narrow, well defined, problem areas. years later a similar survey reported over 2500 developed expert
systems (Durkin, 1994). The new growing area was business
• The DENDRAL project originated the fundamental idea of
and manufacturing, which accounted for about 60% of the
expert systems – knowledge engineering, which encompassed
applications. Expert system technology had clearly matured.
techniques of capturing, analyzing and expressing in rules an
expert’s “know-how”.

23 24
 However:
• Expert systems have difficulty in recognizing domain
• Expert systems are restricted to a very narrow domain of boundaries. When given a task different from the typical
expertise. For example, MYCIN, which was developed for problems, an expert system might attempt to solve it and fail in
the diagnosis of infectious blood diseases, lacks any real rather unpredictable ways.
knowledge of human physiology. If a patient has more than • Heuristic rules represent knowledge in abstract form and lack
one disease, we cannot rely on MYCIN. In fact, therapy even basic understanding of the domain area. It makes the task
prescribed for the blood disease might even be harmful of identifying incorrect, incomplete or inconsistent
because of the other disease. knowledge difficult.
• Expert systems can show the sequence of the rules they • Expert systems, especially the first generation, have little or no
applied to reach a solution, but cannot relate accumulated, ability to learn from their experience. Expert systems are built
heuristic knowledge to any deeper understanding of the individually and cannot be developed fast. Complex systems
problem domain. can take over 30 person-years to build.

25 26

How to make a machine learn, or the rebirth of

neural networks (mid-1980s – onwards)
• By the late sixties, most of the basic ideas and concepts
necessary for neural computing had already been formulated.
However, only in the mid-eighties did the solution emerge. The
• In the mid-eighties, researchers, engineers and experts found major reason for the delay was technological: there were no
that building an expert system required much more than just PCs or powerful workstations to model and experiment with
buying a reasoning system or expert system shell and putting artificial neural networks.
enough rules in it. Disillusions about the applicability of
• In the eighties, because of the need for brain-like information
expert system technology even led to people predicting an AI
processing, as well as the advances in computer technology and
“winter” with severely squeezed funding for AI projects. AI
progress in neuroscience, the field of neural networks
researchers decided to have a new look at neural networks.
experienced a dramatic resurgence. Major contributions to both
theory and design were made on several fronts.

27 28
 Neural networks to Deep Learning

• Grossberg established a new principle of self-organisation

(adaptive resonance theory), which provided the basis for a
new class of neural networks (Grossberg, 1980).
• Hopfield introduced neural networks with feedback –
Hopfield networks, which attracted much attention in the
eighties (Hopfield, 1982).
• Kohonen published a paper on self-organising maps
(Kohonen, 1982).
• Barto, Sutton and Anderson published their work on
reinforcement learning and its application in control (Barto
et al., 1983).

29 30

4. So What Is Machine Learning?

• Getting computers to program themselves

• Writing software is the bottleneck
• Let the data do the work instead!

31 32
 Magic?
Traditional Programming

Data No, more like gardening

Computer Output
Program
◼ Seeds = Algorithms
◼ Nutrients = Data
Machine Learning ◼ Gardener = You
◼ Plants = Programs
Data
Computer Program
Output

33 34

Sample Applications ML in a Nutshell

• Web search • Tens of thousands of machine learning algorithms

• Computational biology • Hundreds new every year
• Finance • Every machine learning algorithm has three components:
• E-commerce • Representation
• Space exploration • Evaluation
• Robotics • Optimization
• Information extraction
• Social networks
• Debugging
• [Your favorite area]
35 36
Representation Evaluation

• Decision trees • Accuracy

• Sets of rules / Logic programs • Precision and recall
• Instances • Squared error
• Graphical models (Bayes/Markov nets) • Likelihood
• Posterior probability
• Neural networks
• Cost / Utility
• Support vector machines
• Margin
• Model ensembles • Entropy
• Etc. • K-L divergence
• Etc.

37 38

Optimization Types of Learning

• Supervised (inductive) learning

• Combinatorial optimization
• Training data includes desired outputs
• E.g.: Greedy search
• Unsupervised learning
• Convex optimization
• Training data does not include desired outputs
• E.g.: Gradient descent
• Semi-supervised learning
• Constrained optimization
• Training data includes a few desired outputs
• E.g.: Linear programming
• Reinforcement learning
• Rewards from sequence of actions

39 40