Lecture 1

◼ What is Artificial Intelligence?

1
Key Questions
◼ Is it possible to simulate intelligent behavior
on computers?
◼ Which are the criteria based on which the
intelligence of a program is evaluated?
◼ At which level the intelligent behavior tries
to be modelled ?
◼ Which are the representations and the
techniques used when solving AI problems?

2
What is AI?
◼ Alan Turing - “Computing Machinery and
Intelligence”, 1950
◼ AI is the study of ideas which allow
computers to become intelligent

3
4
AI Definition
◼ AI studies the creation of computer
systems and programs, which exhibit a
form of intelligence
◼ systems which can learn new concepts
◼ systems which can reason and deduce useful
concepts in a domain
◼ systems which can understand the natural
language or can perceive and process an image
◼ briefly systems which need intelligent abilities,
specific to humans

5
AI Definition
◼ An intelligent program is a program which
exhibit a similar behavior with a human,
when it is faced with a similar problem
◼ It is not necessary for the program to solve
or to try to solve the problem in the same
way a human will do it

6
AI Problems Nature
◼ The problems which must be solved have a
common characteristic
◼ They are difficult problems, NP-complete
◼ Even using a powerful computer, this can
not solve the problem, with an O(eN) time
complexity algorithm, for a huge input
value N, in a reasonable time interval

7
AI Problems Nature
◼ The researches in AI are directed towards
trying to decrease the complexity necessary
for searching the solution of NP-complete
problems
◼ In AI, heuristic functions are used, for
solving difficult problems
◼ In this way, intractable problems, from the
point of view of solving time, can become
possible to be solved, in many cases
8
AI Problems Nature
◼ Another research direction in AI regards
solving trivial problems, which imply
common sense knowledge
◼ These problems include reasoning about
physical objects and their relationship, and
reasoning about actions and their
consequences

9
AI Problems Nature
◼ Everybody knows that:
◼ an object can not be in two places
simultaneously
◼ a glass must not be dropped on the floor,
because it will break
◼ These behaviors are hardly characterized as
needing intelligence
◼ These behaviors are very difficult to be
modelled in a program
10
 AI Problems Nature
◼ Common sense knowledge can be used by
every human, but these must be explicitly
represented in a program, and their quantity is
impressive
◼ Surprisingly, the researches in AI had better
results is domains like solving difficult
problems, such as games or theorem proving,
or problems which need human expertise in a
certain domain, than in domains which need
common sense knowledge
11
AI Problems Nature
◼ Many researches in AI are directed to
mimic human capacities, like speaking,
seeing, hearing
◼ Image perception and recognition, natural
language understanding, natural language
synthesis are very challenging problems,
which imply complicated processing
algorithms and complex technical devices

12
Turing Test
◼ The most famous criterion to estimate a
program intelligence is the Turing test
◼ In 1950, Alan Turing (1912-1954), famous
British mathematician, one of the founders
of computer science, proposed a test to
evaluate if a machine can have an intelligent
behavior
◼ Turing stated for the first time the
possibility of human thinking simulation
problem, using a computer

13
Turing Test
◼ Turing describes the test which can
distinguish between an intelligent and a
non-intelligent behavior
◼ A human communicates through a computer
terminal with two other terminals, placed in
neighboring rooms, by asking questions and
receiving answers

14
Turing Test
◼ The answers are given by a human at one of
the two terminals and by a computer
program at the second terminal
◼ Both terminals are hidden to the person
which asks questions
◼ If the person who asks questions can not
establish, after the dialogue, where is the
terminal from where the human answers
and where is the terminal from where the
computer program answers, then the
program has an intelligent behavior
15
 Turing Test
◼ The program can be made to try to cheat the
person who asks questions
◼ If the program is asked a complex mathematical
computation, then several minutes pass until the
answer is received
◼ The most important problem raised by making
such a program to pass the test is the quantity of
information needed in the program, considering
that the questions can refer to different domains,
and the fact that the questions are asked in
natural language
16
Turing Test
◼ A program for playing chess can have
similar performances to those of a master in
playing chess, but the opponent will never
know if he/she is playing chess with a
person or with a computer program
◼ In other domains, it is possible to compare
the performances of solving a problem by a
computer program to those of a human
expert, from the point of view of the
solution quality and of the solving speed
17
Brief History
◼ Dartmouth College Conference in 1956 –
the first four promoters of the domain: John
McCarthy, Marvin Minsky, Alen Newell,
and Herbert Simon
◼ 1956 - 1957 A. Newell, J. Shaw, and H.
Simon – the first automated theorem prover
- "The Logic Theorist"
◼ Starting from 1960, the first AI programs
appear
18
Brief History
◼ 1965 J. A. Robinson – Resolution principle
◼ 1965 – DENDRAL - J. Lederberg and E.
Feigenbaum - expert system capable to synthetize
organic molecules structure, based on chemical
formulas and mass spectograms
◼ 1959 – Lisp Language (LISt Processing) - John
McCarthy (Dartmouth College)
◼ 1972 – Prolog Language (PROgrammation en
LOGique) - Alain Colmerauer (Marseille
University)
◼ 1983 – Smalltalk - Goldberg, Robson
19
Brief History
◼ 1970 – knowledge is important
◼ Knowledge Based Systems
◼ Knowledge Engineering
◼ MYCIN System - Buchanan, Shortliffe -
expert system used to diagnose blood
bacteria infection, Stanford University -
1974-1975
◼ Expert Systems

20
Brief History
◼ 1990-2000 Deep Learning
◼ 1989 Convolutional Neural Networks with
backpropagation (Yann LeCun)
◼ 1997 LSTM (Long Short-Term Memory)
networks (Sepp Hochreiter and Jurgen
Schmidhuber)
◼ 2006 Deep Belief Networks (Geoffrey
Hinton and Ruslan Salakhutdinov)
◼ 2008 GPU usage for Deep Neural Networks
21
Brief History
◼ 1997 - the supercomputer Deep Blue
defeated chess champion Gary Kasparov

22
Brief History
◼ 2016 - AlphaGo (DeepMind) beat a
professional Go champion (Lee Sedol)

23
AI Nowadays
◼ AI is everywhere
◼ Communications
◼ Process management
◼ Car driving
◼ Medicine
◼ Web applications
◼ Monitoring
◼ Rescue operations

24
25
Robots

26
Assistive Robots

27
Autonomous Cars

28
Deep Learning Architectures

29
30
ChatGPT
◼ Chat Generative Pre-trained Transformer
◼ Large language model-based chatbot
developed by OpenAI and launched in
November 2022
◼ Enable users to refine and steer a
conversation towards a desired length,
format, style, level of detail, and language

31
ChatGPT
◼ Based on the transformer architecture
developed by Google
◼ Fine-tuned for conversational applications
using a combination of supervised and
reinforcement learning techniques
◼ The fine-tuning process leveraged both
supervised learning as well as
reinforcement learning in a process called
reinforcement learning from human
feedback 32
33
Prompt – Instructions for ChatGPT
◼ Prompt - phrase used to give a specific
direction and framework to an AI model,
like ChatGPT
◼ Establishes a scenario or a role that the
AI ​must place itself into to respond in an
appropriate manner to the specified context
◼ This technique of "setting the scene" helps
create a more relevant and focused
interaction, improving the quality and
accuracy of AI responses 34
Characteristics of a Good Prompt
◼ Clarity and context
◼ Concision
◼ Relevance
◼ Give examples
◼ Give explicit instructions
◼ Iterative questions and answers

35
Latent Space
◼ ChatGPT extracts to a numeric latent space
(a set of numbers) the way people put words
into sentences and phrases to communicate
with each other
◼ Capture the essential features of human
communication in a numeric latent space

36
37
What is NOT ChatGPT?
◼ ChatGPT is NOT a database
◼ ChatGPT is NOT a self awareness entity

38
How ChatGPT is trained?
◼ To train a language model - ChatGPT is a
large language model (LLM) - extract all
the text from Wikipedia, break it into small
pieces, and then present to the network (to
ChatGPT) one piece of text at a time, from
which a word is removed, word that is
asked the network to predict it (output it)

39
How ChatGPT is trained?
◼ After the network has predicted, you also
show it the correct word (the one that exists
in Wikipedia)
◼ In this way, the network is able to modify its
defining set of numbers (the latent space),
such that to capture as best as possible the
essence of reality (the set of features)

40
BERT Language Model
◼ Bidirectional Encoder Representations from
Transformers
◼ Open-source machine learning framework
for natural language processing
◼ Designed to help computers understand the
meaning of ambiguous language in text by
using surrounding text to establish context
◼ BERT framework was pre-trained using text
from Wikipedia and can be fine-tuned with
question-and-answer datasets 41
Ambiental Intelligence (AmI)
◼ Multitouch screens are integrated in
different surfaces – walls, floors, tables
◼ Multitoe – a floor interface, controlled by
the users' feet

42
 AmI

◼ Sixth Sense Project

43
Internet of Things
◼ Make the most surrounding objects to be
interconnected, to function efficiently, but
also to learn and to adapt to the user
behavior
◼ It was identified and defined in 1999 by
Kevin Ashton, a British scientist

44
Internet of Things
◼ In 2010, 4 billions objects from all over the
world were connected to the Internet
◼ In 2012, 15 billions objects were connected
◼ In 2022, 80 billions objects from all over
the world were connected to the Internet

45
Internet of Things
◼ Persons using Nike+ FuelBand to monitor
the distance and the calories are already
using the Internet of Things
◼ In the United States, the homes already
have 10 interconnected devices, and the
number is estimated to increase up to 50 by
2025

46
Internet of Things Scenarios
◼ An intelligent fridge, connected to the
Internet, which knows its content, retains
the foods you like and announces when the
milk is about to finish
◼ Moreover, it can order online the missing
things from the fridge

47
Internet of Things Scenarios
◼ You forget the gas turned on ?
◼ Open a smartphone application and you can
turn it off, no matter where you are
◼ It is possible to make settings such that,
every time when you go out or you come
home, certain operations are automatically
performed
◼ Turn the lights on/off
◼ Control the heating system
48
Internet of Things Scenarios
◼ Intelligent traffic lights, which adapt
themselves the time of green and red lights,
according to real time traffic
◼ Intelligent traffic lights are connected to the
car processor, to block the car to pass the
crossroad on the red light
◼ The car follows the carriage way, because
small sensors from the road marker painting
signal any dangerous deviation
49
Internet of Things Scenarios
◼ Intelligent City – the communication
between all its systems (transport, water,
drainage, electricity, lighting) ensures
energy efficiency, pollution reduction, and
citizens comfort augmentation
◼ South-Korean city Songdo is the first
“smart city”, in which the Internet of Things
is ubiquitous

50
Observations
◼ Internet of Things presume the information
circulation from one point to another and its
usage for actual actions
◼ While the technology will advance and the
surrounding things intelligence will increase
by connectivity, it will be possible to think
at the scenario in which we can search on
Google the objects from the house

51
AI Usage
◼ Generative Art - AI to create digital art,
from paintings to abstract graphics
◼ Chat - AI can help to automate chats with
customers or create chatbots for various
services
◼ Copywriting - AI can help to write texts for
advertisements, blog articles or website
content

52
AI Usage
◼ Translations - AI to translate documents or
real-time conversations in various
languages
◼ Video Editing - automate some parts of the
video editing, such as trimming clips or
adding effects
◼ Music - AI ​can compose music or help you
to develop new songs

53
AI Usage
◼ Vocal Modulation - modifies and adapts
voices in audio recordings, useful for
podcasts or games
◼ Generative Video - creates videos starting
from textual descriptions or turn scenarios
into animations

54
AI Usage
◼ Productivity - improves the efficiency at
work, automating repetitive tasks or
organizing better the information
◼ Research - AI to collect and analyze data,
which can help you make new discoveries
or streamline research processes

55