Artificial intelligence

Chapter 1: What is AI?

I. How should we define AI?
AI is about artificial life-forms that can surpass human intelligence, but it is a very abstract
concept. To frame this, we highlight 3 applications.
Application 1. Self-driving cars
These cars need a combination of AI techniques of many kinds. As a result of this intelligence
that might surpass human level, road safety should eventually improve.
Application 2. Content recommendation
Personalized advertisements, recommendations on Spotify or other social media platforms.
This is in contrast to newspapers, magazines,… where the frontpage is the same for everyone
and isn’t based on artificial algorithms.
Application 3. Image and video processing
Example for this topic is face recognition, similar techniques can be used to recognize other
objects, persons,…
It can also be used to generate or alter visual content like computer generated characters in
animation movies.
The reason why the public perception of AI is so nebulous is first of all because there is no
officially agreed definition. Further there is the legacy of science fiction and last there is that
AI is seen as something easy but is actually really hard and that things what seems hard like
solving mathematical exercises is actually easy.
Key terminology:
1) Autonomy: The ability to perform tasks in complex environments without constant
guidance by a user.
2) Adaptivity: The ability to improve performance by learning from experience.
General definition for AI: the ability of a digital computer or computer-controlled robot to
perform tasks commonly associated with intelligent beings.

II. Related fields

There are other closely related topics such as machine learning, data science and deep
learning.
Machine learning: systems that improve their performance in a given task with more and
more experience or data (example of adaptivity).
Deep learning: a type of machine learning based on artificial neural networks in which
multiple layers of processing are used to extract progressively higher level features from data.
The goal is to increase the amount of knowledge and produce new knowledge of a certain
topic.
Data Science: an umbrella term that includes machine learning and statistics, certain aspects
of computer science including algorithms, data storage, and web application development.
Robotics: building and programming robots so that they can operate in complex, real-world
scenarios. It requires a combination of virtually all areas of AI like computer vision and
speech recognition, natural language processing, information retrieval,…

III. Philosophy of AI
An entity is intelligent if it cannot be distinguished from another intelligent entity by
observing its behavior, but does being human-like mean you are intelligent? The word
intelligent can also be replaced by conscious, because a machine doesn’t has a mind in the
way that a human has.
Is a self-driving car intelligent? The AI system in the car doesn’t see or understand its
environment, and it doesn’t know how to drive safely, in the way a human being sees,
understands, and knows. This means that the intelligent behavior of the system is
fundamentally different from actually being intelligent.
How much does philosophy matter in practice? The philosophy of AI is unlikely to have more
effect on the practice of AI research than philosophy of science generally has on the practice
of science. We continue investigating systems without asking too much whether they are
intelligent or just behave.

Key terminology:
1) General vs narrow AI: narrow refers to AI that handles one task, general AI (AGI) refers to
a machine that can handle any intellectual task. All the AI systems we use now are narrow,
while AGI is considered as science fiction.
2) Strong vs weak AI: strong: intelligent and self-conscious, weak: intelligent behavior,
despite being just computers.
Chapter 2: AI problem solving
I. Search and problem solving
There are many ways to solve a problem, some of them which may be more preferable in
terms of time, effort, cost or other criteria. Different search techniques may lead to different
solutions resulting in advanced search algorithms.
Two kinds of problems:
- Search and planning in static environments with only one “agent”
- Games with two-player (“agents”) competing against each other
First thing to do for the problem solving process = defining the choices and their
consequences. We also need to define what the goal is.
Key terminology:
1) The state space: the set of possible solutions, if the task navigate from place A to B, the
state space could be set of locations defined by their (x,y) coordinates that can be reached
from the starting point.
2) Transitions: are possible moves between one state and another. We only count direct
transitions that can be accomplished with a single action as transition. A sequence of multiple
transitions that follow each other, is called a path.
3) Costs: refer to the fact that, oftentimes the different transitions aren’t all alike. They can
differ in ways that make some transitions more preferable or cheaper and other more costly
(figuratively). If all transitions are equal, we can ignore the costs.

II. Solving problems with AI

Any element of intelligence can be broken down into small steps so that each of the steps is as
such so simple that it can be written down as a computer program.
The most distinctive AI problems in the beginning were games, they provided a convenient
restricted domain that could be formalized easily.
Closely related to games, search and planning techniques were an area where AI led to great
advance: algorithms who were developed back in the 1960s are still the basis for game
playing AI, although more advanced variants have been proposed.

III. Search and games

Example: tic-tac-toe; beginning position of the game= empty grid without Xs or Os played
yet. On the second level, there are possible states that can result from the first player’s move,
be it X or O. These nodes are “children” of the root note (the beginning position). Each nodes
have its children notes and soon on until the game has ended and a player gets a line of three
and wins, or the board is full and the game ends in a tie.
Sample game tree (in the middle of a game):
For the winning position of O, or positions who leads to the victory of On we add the value -
1, the same for X but then +1, if it leads to a tie the value will be 0.

The minmax algorithm guarantees optimal game play in, theoretically speaking, any
deterministic, two-person, perfect information zero-sum game. The algorithm can be
implemented using a few lines of code.
Heuristic evaluation function takes as input a board position, including the information about
which player’s turn is next, and returns a score that should be an estimate of the likely
outcome of the game continuing from the given board position. Good heuristics for chess for
example is typically count the amount of material (pieces) weighted by their type.

Chapter 3: Real world AI

I. Odds and probability
In the real world there is a host of unknown possibilities instead of perfect information as we
discussed in chapter 2, these are ranging from missing information to deliberate deception.
We use the example of a self-driving car, the goal of the car can be set to drive from point A
to point B, but what happens if the traffic get worse, or there is bad weather,… This is why a
self-driving car need to use a variety of sensors, including cameras, to detect where it is and
what is around it.
PROBABILITY
One of the reasons why modern AI methods actually work in real-world problems is their
ability to deal with uncertainty. But the most AI based applications that we use now are based
on probability, as it has turned out to be the best approach for reasoning under uncertainty.
We are most familiar with applications of probability in games: what are the chances of
winning in the lottery or what are the chances to get a royal flush in poker,… However, far
more importantly, probability can also be used to quantify and compare risks in everyday life:
what are the chances of crashing your car if you exceed the speed limit, or what are the
chances that the interest rates on your loan will go up by 5%,…
Probability can be quantified (expressed as a number) and it can be right or wrong.
ODDS
Odds make it particularly easy to update beliefs when more information becomes available.
By odds, we mean an expression like 3:1, which means that we expect that for every three
cases of an outcome, for example winning a bet.
The odds 1:5, is different from 20% probability, 1:5 mean that you’d have to play a game six
times to get one win on the average. The probability 20% means that you have to play five
times to get one win on the average.

II. The Bayes rule

Key terminology:
1) Prior and posterior odds: the simplest form to express the bayes rule is in terms of odds.
Prior odds are the odds for something to happen (against it not happening), posterior odds are
updated prior odds with new information.
2)Likelihood ratio: is the probability of the observation in case the event of interest, divided
by the probability of the observation of no event. For example if the chance that it will rain is
9/10, the likelihood ratio= (9/10)/(1/10)=9.
Bayes Rule for converting prior odds into posterior odds
-> Posterior odds= likelihood ratio x prior odds

III. Naive Bayes classification

The Bayes classifier is a machine learning technique that can be used to classify objects such
as text documents into two or more classes. It is trained by analyzing a set of training data, for
which correct classes are given. Example: spam filters.

Chapter 4: Machine learning

I. The types of machine learning
1) Supervised learning: given an input and the task is to predict the correct output or label, in
the simplest cases, the answers are in the form yes/no (binary classification problems.
2) Unsupervised learning: there are no labels or correct outputs. The task is to discover the
structure of the data. Example: data visualization.
3) Reinforcement learning: commonly used in situations where an AI agent must operate in an
environment and where feedback about good or bad choices is available with some delay
(learning from mistakes).

II. The nearest neighbor classifier

In the image above, we use the geometric distance (Euclidean distance) to classify the input.
However, in some cases it might not be clear how the items are drawn in a geometric
representation and how distances should be measured. You should therefore choose the
distance metric on a case-by-case basis.

III. Regression
Linear regression
What is linear regression? AI Linear Regression is a supervised machine learning algorithm
with a continuous and constant slope expected performance. Rather than attempting to
classify values into groups (such as cats and dogs), it is used to estimate values within a
continuous range (such as price and revenue).
Linear regression produces a numerical prediction that is not constrained to be an integer, it is
better suited in situations where the output variable can be any number like the price of a
product, the distance to an obstacle,…
In linear regression, the main idea is to add up the effects of each of the feature variables to
produce the predicted value. The term of the process is called linear combination.
Example of linear regression as a shopping bill:
Suppose you go to the grocery store and buy 2.5kg potatoes, 1.0kg carrots, and two bottles of
milk. If the price of potatoes is 2€ per kg, the price of carrots is 4€ per kg, and a bottle of milk
costs 3€, then the bill, calculated by the cashier, totals 2.5 × 2€ + 1.0 × 4€ + 2 × 3€ = 15€. In
linear regression, the amount of potatoes, carrots, and milk are the inputs in the data. The
output is the cost of your shopping, which clearly depends on both the price and how much of
each product you buy.
Linear means that the increase in the output when one input feature is increased by some fixed
amount is always the same.
Key terminology:
Coefficients or weights: This are the prices of the different products, one of the main
advantages of linear regression is its easy interpretability: the learned weight may in fact be
more interesting than the predictions of the outputs.
The starting point for a calculation in linear regression is the intercept.
Some examples of applications of linear regression are: prediction of click rates in online
advertising, prediction of retail demand for products, prediction of software cost, prediction of
insurance cost, prediction of crime rates,…

Logistic regression
In logistic regression, we can turn the linear regression method’s outputs into predictions
about labels. We take the output from linear regression, which is a number, and predict one
label A if the output is greater than zero, and another label B if the output is less than or equal
to zero. Instead of just predicting one class or another, logistic regression can also give us a
measure of uncertainty of the prediction. It is also possible to use it to obtain predictions over
more than two possible labels.
Chapter 5: Neural networks
I. Neural network basics
A neural network can mean either a “real” biological network such as the one in our brain, or
an artificial neural network simulated in a computer.
Key terminology:
1)Deep learning refers to certain kinds of machine learning techniques where several “layers”
of simple processing units are connected in a network so that the input to the system is passed
through each one of them in turn.
2) Neurons, cell bodies and signals: A neural network consists of a large number of simple
units, neurons, that receive and transmit signals to each other. The neurons are very simple
processors of information, consisting of a cell body and wires that connect the neurons to each
other.
3) Dendrites, axons and synapses: Dendrites are wires that provide the input to the neurons.
The wire that transmits the outgoing signal is called an axon. Each axon may be connected to
one or more dendrites at intersections that are called synapses.
II. How neural networks are built
Once the linear combination has been computed, the neuron takes the linear combination and
puts it through a so-called activation function. Examples: identify function (do nothing and
just output the linear combination), step function (if the value of the linear combination is
greater than zero, send a pulse ON, otherwise do nothing OFF), sigmoid function (a “soft”
version of the step function).
The output of the neuron, determined by the linear combination and the activation function,
can be used to extract a prediction or a decision.
Key terminology
Layers: Often the network architecture is composed of layers. The input layer consists of
neurons that get their inputs directly from the data. Hidden layers use the other neurons’
outputs as their input, and whose output is used as the input to other layers of neurons. The
output layer produces the output of the whole network. All the neurons on a given layer gets
inputs from neurons on the previous layer and feed their output to the next.

III. Advanced neural network techniques

Convolutional neural networks (CNNs): this is a solution of the weights becoming extremely
large and consequently, and the amount of training data required to achieve satisfactory
accuracy become too large to be realistic. This special kind of neural network has a special
kind of layer (convolutional layer). Networks including convolutional layers are called
convolutional neural networks (CNNs). Their key property is that they can detect image
features such as bright or dark (or specific color) spot, edges in various orientations, patterns,
and so on. It would normally be hard to train a neural network for these difficult tasks,
because the features can appear in different positions, orientations, sizes,… CNN can
recognize the object anywhere in the image no matter where it has been observed in the
training images. The trick that CNN uses is using the same input weights for many neurons –
so that all of these neurons are activated by the same pattern – but with different input pixels.
Generative adversarial networks (GANs): The idea is to let two networks compete against
each other. One of the networks is trained to generate images like the ones in the training data,
while the other network’s task is to separate images generated by the first network from real
images from the training data.

Chapter 6: Implications
I. About predicting the future
Predicting the future of AI is nearly impossible, we don’t know what impact AI will have on
our life in the future. However, by understanding the history and present of AI we might be
able to set some predictions.
Prediction 1: AI will continue to be all around us; There will most likely be some applications
of AI in the future such as self-driving cars, recommendation systems, and image and video
processing.
Prediction 2: the Terminator isn’t coming; In this scenario, AI will continue to grow, but there
will robots with advanced killing techniques who wipes most of the humankind out of
existence with nukes.
Prediction 3: Doomsday scenario 1: There will be a powerful AI system who just becomes
conscious and decides that it really dislikes humanity.
Prediction 4: There will be a robot army controlled by an intelligent but not conscious AI
system that is in principle in human control.
However, some of these scenarios seem to be really unrealistic.

II. The societal implications of AI

Implication 1: Algorithmic bias; This means the embedding of a tendency to discriminate
according ethnicity, gender, or other factors when making decisions about job applications,
bank loans, and so on. The main reason for algorithmic bias is human bias in the data.
Implication 2: Seeing is believing?; When we see an authority or trustworthy person stating
something, we are most likely tended to trust them. The same thing can be applied for other
cases. We are of course aware of the possibility of fabricating fake evidence, like deep face or
photoshop.
Implication 3: Changing notions of privacy; de-anonymization is breaking the anonymity of
data that we may have thought to be safe. The basic problem is that when we report the results
of an analysis, the results may be so specific that they make it possible to learn about
individual users. There are a lot more applications where companies are able to collect
personal information from users.
Implication 4: Changing work; each step towards better automation changes the working life.
With AI and robotics, there is even less need for many kinds of dull, repetitive work.