A non-technical introduction to machine learning

Olivier Colliot

To cite this version:

Olivier Colliot. A non-technical introduction to machine learning. Olivier Colliot. Machine Learning
for Brain Disorders, Springer, 2023. �hal-03957125v3�

HAL Id: hal-03957125

https://hal.science/hal-03957125v3
Submitted on 24 May 2023

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.

Distributed under a Creative Commons Attribution| 4.0 International License

 Chapter 1

A non-technical introduction
to machine learning

Olivier Colliot*1
1 Sorbonne Université, Institut du Cerveau - Paris Brain Institute - ICM,
CNRS, Inria, Inserm, AP-HP, Hôpital de la Pitié-Salpêtrière, F-75013, Paris,
France
* Corresponding author: e-mail address: [email protected]

Abstract
This chapter provides an introduction to machine learning for a non-
technical readership. Machine learning is an approach to artificial intelli-
gence. The chapter thus starts with a brief history of artificial intelligence
in order to put machine learning into this broader scientific context. We
then describe the main general concepts of machine learning. Readers
with a background in computer science may skip this chapter.

Keywords: machine learning, artificial intelligence, supervised learn-

ing, unsupervised learning

1. Introduction
Machine learning (ML) is a scientific domain which aims at allowing
computers to perform tasks without being explicitly programmed to do
so [1]. To that purpose, the computer is trained using the examination
of examples or experiences. It is part of a broader field of computer
science called artificial intelligence (AI) which aims at creating comput-
ers with abilities that are characteristic of human or animal intelligence.
This includes tasks such as perception (the ability to recognize images
or sounds), reasoning, decision making or creativity. Emblematic tasks

O. Colliot (Ed.), Machine Learning for Brain Disorders, Springer, 2023

2 Colliot

which are easy to perform for a human and are inherently difficult for a
computer are for instance recognizing objects, faces or animals in pho-
tographs, or recognizing words in speech. On the other hand, there are
also tasks which are inherently easy for a computer and difficult for a hu-
man, such as computing with large numbers or memorizing exactly huge
amounts of text. Machine learning is the AI technique that has achieved
the most impressive successes over the past years. However, it is not the
only approach to AI and conceptually different approaches also exist.
Machine learning also has close ties to other scientific fields. First,
it has evident strong links to statistics. Indeed, most machine learning
approaches exploit statistical properties of the data. Moreover, some
classical approaches used in machine learning were actually invented in
statistics (for instance linear or logistic regression). Nowadays, there is
a constant interplay between progress in statistics and machine learn-
ing. ML has also important ties to signal and image processing, ML
techniques being efficient for many applications in these domains and
signal/image processing concepts being often key to the design or un-
derstanding of ML techniques. There are also various links to different
branches of mathematics, including optimization and differential geom-
etry. Besides, some inspiration for the design of ML approaches comes
from the observation of biological cognitive systems, hence the connec-
tions with cognitive science and neuroscience. Finally, the term data
science has become commonplace to refer to the use of statistical and
computational methods for extracting meaningful patterns from data. In
practice, machine learning and data science share many concepts, tech-
niques and tools. Nevertheless, data science puts more emphasis on the
discovery of knowledge from the data while machine learning focuses on
solving tasks.
This chapter starts by providing a few historical landmarks regarding
artificial intelligence and machine learning (section 2). It then proceeds
with the main concepts of ML which are foundational to understand
other chapters of this book.

2. A bit of history
As a scientific endeavour, artificial intelligence is at least 80 years old.
Here, we provide a very brief overview of this history. For more details,
the reader may refer to [2]. A non-exhaustive timeline of AI is shown
in Figure 1.
Even if this is debatable, one often considers AI to emerge in the
1940s–1950s with a series of important concepts and events. In 1943,
the neurophysiologist Warren McCulloch and the logician Walter Pitts

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 3

Turing
Expert
test
systems Deep
learning
SVM era
Dartmouth
Perceptron Back-
workshop
propagation

Artificial
neuron model

1940 1950 1960 1970 1980 1990 2000 2010 2020

Figure 1: A brief timeline of AI with some of the landmark advances.

proposed an artificial neuron model, which is a mathematical abstraction

of a biological neuron [3], and showed that sets of neurons can compute
logical operations. In 1948, the mathematician and philosopher Norbert
Wiener coined the term “cybernetics“ [4] to designate the scientific study
of control and communication in humans, animals and machines. This
idea that such processes can be studied within the same framework in
both humans/animals and machines is a conceptual revolution. In 1949,
the psychologist Donald Hebb [5] described a theory of learning for bi-
ological neurons which was later influential in the modification of the
weights of artificial neurons.
In 1950, Alan Turing, one of the founders of computer science, intro-
duced a test (the famous “Turing test”) for deciding if a machine can
think [6]. Actually, since the question can a machine think? is ill-posed
and depends on the definition of thinking, Turing proposed to replace
it with a practical test. The idea is that of a game in which an inter-
rogator is given the task of determining which of two players A and B
is a computer and which is a human (by using only responses to writ-
ten questions). In 1956, the mathematician John McCarthy organized
what remained as the famous Darmouth workshop and which united
ten prominent scientists for two months (among which Marvin Minsky,
Claude Shannon, Arthur Samuel and others). This workshop is more im-
portant by its scientific programme than by its outputs. Let us reproduce
here the first sentences of the proposal written by McCarthy et al. [7] as
we believe that they are particularly enlightening on the prospects of
artificial intelligence:

We propose that a 2 month, 10 man study of artificial intelli-

gence be carried out during the summer of 1956 at Dartmouth
College in Hanover, New Hampshire. The study is to proceed

Machine Learning for Brain Disorders, Chapter 1

4 Colliot

on the basis of the conjecture that every aspect of learning or

any other feature of intelligence can in principle be so pre-
cisely described that a machine can be made to simulate it.
An attempt will be made to find how to make machines use
language, form abstractions and concepts, solve kinds of prob-
lems now reserved for humans, and improve themselves. We
think that a significant advance can be made in one or more of
these problems if a carefully selected group of scientists work
on it together for a summer.

There was no major advance made at the workshop, although a reasoning

program, able to prove theorems was presented by Allen Newell and
Herbert Simon [8] at this occasion. This can be considered as the start
of symbolic AI (we will come back later on the two main families of AI:
symbolic and connexionist). Let us end the 1950s with the invention,
in 1958, of the perceptron by Frank Rosenblatt [9], whose work was
built upon the ideas of McCulloch, Pitts and Hebb. The perceptron
was the first actual artificial neuron. It was able to recognize images.
This is an important landmark for several reasons. The perceptron, with
some modifications, is still the building block of modern deep learning
algorithms. To mimic an artificial neuron (Figure 2), it is composed of a
set of inputs (which correspond to the information entering the synapses)
xi , which are linearly combined and then go through a non-linear function
g to produce an output y. This was an important advance at the time
but it had strong limitations, in particular its inability to discriminate
patterns which are not linearly separable. More generally, in the field of
AI as a whole, unreasonable promises had been made and they were not
delivered: newspapers were writing about upcoming machines that could
talk, see, write and think; the US government funded huge programs to
design automatic translation programs, etc. This led to a dramatic drop
in research funding and, more generally, in interest in AI. This is often
referred to as the first AI winter (Figure 3).
Even though research in AI continued, it was not before the early
1980s that real-world applications were once again considered possible.
This wave was that of expert systems [10], which are a type of sym-
bolic AI approach but with domain-specific knowledge. Expert systems
led to commercial applications and to a real boom in the industry. A
specific programming language, called LISP [11], became dominant for
the implementation of expert systems. Companies started building LISP
machines, which were dedicated computers with specific architecture tai-
lored to execute LISP efficiently. One cannot help thinking of a parallel
with current hardware dedicated to deep learning. However, once again,
expectations were not met. Expert systems were very large and complex

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 5

(a)
Dendrites Axon terminal

x1
Cell body
x2 y
..
.
xp

Axon

(b)

!! $!

$"
!" ∑ "

$# %
…

!#

Inputs Weights Sum Non-Linearity Output

Figure 2: (a) Biological neuron. The synapses form the input of the
neuron. Their signals are combined and if the result exceeds a given
threshold, the neuron is activated and produces an output signal which is
sent through the axon. (b) The perceptron: an artificial neuron which is
inspired by biology. It is composed of the set of inputs (which correspond
to the information entering the synapses) xi , which are linearly combined
with weights wi and then go through a non-linear function g to produce
an output y. Image in panel (a) is courtesy of Thibault Rolland.

Machine Learning for Brain Disorders, Chapter 1

6 Colliot

Machine
learning
Expert
systems

Early research
2nd AI
1st
AI winter
winter
1940 1950 1960 1970 1980 1990 2000 2010 2020

Figure 3: Summers and winters of AI.

sets of rules. They were difficult to maintain and update. They also had
poor performances in perception tasks such as image and speech recogni-
tion. Academic and industrial funding subsequently dropped. This was
the second AI winter.
At this stage, it is probably useful to come back to the two main
families of AI: symbolic and connexionist (Figure 4). They had important
links at the beginning (see for example the work of McCulloch and Pitt
aiming to perform logical operations using artificial neurons) but they
subsequently developed separately. In short, these two families can be
described as follows. The first operates on symbols through sets of logical
rules. It has strong ties to the domain of predicate logic. Connexionism
aims at training networks of artificial neurons. This is done through
the examination of training examples. More generally, it is acceptable
to put most machine learning methods within the connexionist family,
even though they don’t rely on artificial neuron models, because their
underlying principle is also to exploit statistical similarities in the training
data. For a more detailed perspective on the two families of AI, the reader
can refer to the very interesting (and even entertaining!) paper of Cardon
et al [12].
Let us come back to our historical timeline. The 1980s saw a rebirth of
connexionism and, more generally, the start of the rise of machine learn-
ing. Interestingly, it is at that time that two of the main conferences
on machine learning started: the International Conference on Machine
Learning (ICML) in 1980 and Neural Information Processing Systems
(NeurIPS, formerly NIPS) in 1987. It had been known for a long time
that neural networks with multiple layers (as opposed to the original per-
ceptron with a single layer) (Figure 5) could solve non-linearly separable
problems but their training remained difficult. The back-propagation al-

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 7

Artificial intelligence

Symbolic AI Connexionism
Operates on symbols - Neural networks
through logical rules - More generally:
machine learning

Figure 4: Two families of AI. The symbolic approach operates on sym-

bols through logical rules. The connexionist family actually not only en-
compasses artificial neural networks but more generally machine learning
approaches.

gorithm for training multilayer neural networks was described by David

Rumelhart, Geoffrey Hinton and Ronald Williams [13] in 1986, as well
as by Yann Le Cun in 1985 [14], who also refined the procedure in his
PhD thesis published in 1987. This idea had actually been explored
since the 1960s but it was only in the 1980s that it was efficiently used
for training multilayer neural networks. Finally, in 1989, Yann Le Cun
proposed the convolutional neural network [15], an architecture inspired
by the organization of the visual cortex, whose principle is still at the
core of state-of-the-art algorithms for many image processing and recog-
nition tasks. Multilayer neural networks demonstrated their utility in
several real-world applications such as digit recognition on checks and
ZIP codes [16]. Nevertheless, they would not become the dominant ma-
chine learning approach until the 2010s. Indeed, at the time, they re-
quired considerable computing power for training and there was often
not enough training data.
During the 1980s and 1990s, machine learning methods continued to
develop. Interestingly, connections between machine learning and statis-
tics increased. We are not going to provide an overview of the history
of statistics, but one should note that many statistical methods such
as linear regression (Legendre, 1805 [17]), principal component analysis
(Pearson, 1901 [18]), discriminant analysis (Fisher, 1936) [19] or decision
trees [20] can actually be used to solve machine learning tasks such as
automatic categorization of objects or prediction. In the 1980s, decision
trees witnessed important developments (see for example the ID3 [21] and
CART [21] algorithms). In the 1990s, there were important advances in

Machine Learning for Brain Disorders, Chapter 1

8 Colliot

&!

!!

&" "'!

!"

&$ "'"
!#

&%!

Inputs Hidden Outputs

Figure 5: A multilayer perceptron model (here with only one hidden

layer, but there can be many more).

the statistical theory of learning (in particular the works of Vladimir

Vapnik [22]). A landmark algorithm developed at that time was the
support-vector machine (SVM) [23] which worked well with small train-
ing data sets and could handle non-linearities through the use of kernels.
The machine learning field continued to expand through the 2000s and
2010s, with new approaches but also more mature software packages such
as scikit-learn [24]. More generally, it is actually important to have in
mind that what is currently called AI owes more to statistics (and other
mathematical fields such as optimization in particular) than to modeling
of brain circuitry and that even approaches that take inspiration from
neurobiology can actually be viewed as complex statistical machineries.
2012 saw the revival of neural networks and the beginning of the era of
deep learning. It was undoubtedly propelled by the considerable improve-
ment obtained on the ImageNet recognition challenge which contains 14
million natural images belonging to 20,000 categories. The solution, pro-
posed by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton [25], was
a convolutional neural network with a large number of layers, hence the

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 9

term deep learning. The building blocks of this solution were already
present in the 1980s but there was not enough computing power nor
large training datasets for them to work properly. In the interval, things
had changed. Computers had become exponentially more powerful and,
in particular, the use of Graphical Processing Units (GPU) considerably
sped up computations. The expansion of the Internet had provided mas-
sive amounts of data of various sorts such as texts and images. In the
subsequent years, deep learning [26] approaches became increasingly so-
phisticated. In parallel, efficient and mature software packages including
TensorFlow [27], PyTorch [28] or Keras [29], whose development is sup-
ported by major companies such as Google and Facebook, enable deep
learning to be used more easily by scientists and engineers.
Artificial intelligence in medicine as a research field is about 50 years
old. In 1975, an expert system, called MYCIN, was proposed to identify
bacteria causing various infectious diseases [30]. More generally, there
was a growing interest in expert systems for medical applications. Med-
ical image processing also quickly became a growing field. The first con-
ference on Information Processing in Medical Imaging (IPMI) was held
in 1977 (it existed under a different name since 1969). The first SPIE
Medical Image Processing conference took place in 1986 and the Med-
ical Image Computing and Computer-Assisted Intervention conference
(MICCAI) started in 1998. Image perception tasks, such as segmenta-
tion or classification, soon became among the key topics of this field,
even though the methods came in majority from traditional image pro-
cessing and not from machine learning. In the 2010s, machine learning
approaches became dominant for medical image processing and more
generally in artificial intelligence in medicine.
To conclude this part, it is important to be clear about the different
terms, in particular those of artificial intelligence, machine learning and
deep learning (Figure 6). Machine learning is one approach to artificial
intelligence and other radically different approaches exist. Deep learning
is a specific type of machine learning approach. It has recently obtained
impressive results on some types of data (in particular images and text)
but this does not mean that it is the universal solution to all problems.
As we will see in this book, there are tasks for which other types of
approaches perform best.

3. Main machine learning concepts

As aforementioned, machine learning aims at making a computer capable
of performing a task without explicitly being programmed for that task.
More precisely, it means that one will not write a sequence of instruc-

Machine Learning for Brain Disorders, Chapter 1

10 Colliot

Artificial intelligence

Machine learning

Deep learning

Figure 6: Artificial intelligence, machine learning and deep learning

are not synonymous. Deep learning is a type of machine learning which
involves neural networks with a large number of hidden layers. Machine
learning is one approach to artificial intelligence, but other approaches
exist.

tions that will directly perform the considered task. Instead, one will
write a program that allows the computer to learn how to perform the
task by examining examples or experiences. The output of this learning
process is a computer program itself that performs the desired task, but
this program was not explicitly written. Instead, it has been learned
automatically by the computer.
In 1997, Tom Mitchell gave a more precise definition of a well-posed
machine learning problem [31]:

A computer program is said to learn from experience E with

respect to some task T and some performance measure
P, if its performance at task T, as measured by P, improves
with experience E.

He then provides the example of a computer that learns to play checkers:

task T is playing checkers, performance measure P is the proportion of
games won and the training experience E is playing checker games against
itself. Very often, the experience E will not be an actual action but the
observation of a set of examples, for instance a set of images belonging
to different categories such as photographs of cats and dogs, or medical
images containing tumours or without lesions. Please refer to Box 1 for
a summary.

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 11

Box 1: Definition of machine learning

Machine learning definition (Mitchell, 1997 [31]):

• “a computer program is said to learn from experience E with

respect to some task T and some performance measure P,
if its performance at task T, as measured by P, improves with
experience E.”

Example: learning to detect tumours from medical images

• Task T: detect tumours from medical image

• Performance measure P: proportion of tumours correctly

identified

• Experience E: examining a dataset of medical images where

the presence of tumours has been annotated

3.1 Types of learning

One usually considers three main types of learning: supervised learn-
ing, unsupervised learning and reinforcement learning (Box 2). In both
supervised and unsupervised learning, the experience E is actually the
inspection of a set of examples, which will we refer to as training examples
or training set.

3.1.1. Supervised learning

In supervised learning, the machine learns to perform a task by examin-

ing a set of examples for which the output is known (i.e. the examples
have been labelled). The two most common tasks in supervised learn-
ing are classification and regression (Figure 7). Classification aims at
assigning a category for each sample. The examples can for instance
be different patients and the categories are the different possible diag-
noses. The outputs are thus discrete. Examples of common classification
algorithms include logistic regression (in spite of its name, it is a classi-
fication method), linear discriminant analysis, support vector machines,
random forest classifiers and deep learning models for classification. In
regression, the output is a continuous number. This can be for example
the future clinical score of a patient that we are trying to predict. Ex-

Machine Learning for Brain Disorders, Chapter 1

12 Colliot

Box 2: Supervised, unsupervised and reinforcement learning

• Supervised learning. Learns from labelled examples, i.e.

examples for which the output that we are trying to learn is
known.

– Example 1. The task is computer-aided diagnosis (a clas-

sification problem), the label can be the diagnosis of each
patient, as defined by an expert physician.
– Example 2. The task is the prediction of the age of a
person from a set of biological variables (for example a
brain MRI). This is a regression problem. The label is
the true age of a given person in the training set.

• Unsupervised learning. Learns from unlabelled examples.

– Example 1. Given a large set of newspapers articles, auto-

matically cluster them into groups dealing with the same
topic based only on the text of the article. The topics can
for example be economics, politics, international affairs.
The topics are not known a priori.
– Example 2. Given a set of patients with autism spec-
trum disorders, the aim is to discover cluster of patients
that share the same characteristics. The clusters are not
known a priori. Examples 1 and 2 will be referred to as
clustering tasks.
– Example 3. Given a large set of medical characteristics
(various biological measurements, clinical and cognitive
tests, medical images), find a small set of variables that
best explain the variability of the dataset. This is a di-
mensionality reduction problem.

• Reinforcement learning. Learns by iteratively performing

actions to maximize some reward.

– Classical approach used for learning to play games (chess,

go. . . ) or in the domain of robotics.
– Currently few applications in the domain of brain dis-
eases.

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 13

Classification

Regression
30.0

27.5
Body mass index

25.0

22.5

20.0

17.5

20 30 40 50 60
Age (in years)

Figure 7: Two of the main supervised learning tasks: classification and

regression. The upper panel presents a classification task which aims at
linearly separating the orange and the blue class. Each sample is described
by two variables. The lower panel presents a linear regression task in
which the aim is to predict the body mass index from the age of a person.
Figure courtesy of Johann Faouzi.

Machine Learning for Brain Disorders, Chapter 1

14 Colliot

amples of common regression methods include simple or multiple linear

regression, penalized regression, and random forest regression. Finally,
there are many other tasks that can be framed as a supervised learning
problem, including for example data synthesis, image segmentation and
many others which will be described in other chapters of this book.

3.1.2. Unsupervised learning

In unsupervised learning, the examples are not labeled. The two most
common tasks in unsupervised learning are clustering and dimensionality
reduction (Figure 8). Clustering aims at discovering groups within the
training set but these groups are not known a priori. The objective is
to find groups such that members of the same group are similar while
members of different groups are dissimilar. For example, one can aim
to discover disease subtypes which are not known a priori. Some clas-
sical clustering methods are k-means or spectral clustering for instance.
Dimensionality reduction aims at finding a space of variables (of lower
dimension than the input space) that best explain the variability of the
training data, given a larger set of input variables. This produces a new
set of variables that, in general, are not among the input variables but
are combinations of them. Examples of such methods include principal
component analysis, Laplacian eigenmaps or variational autoencoders.

3.1.3. Reinforcement learning

In reinforcement learning, the machine will take a series of actions in
order to maximize a reward. This can for example be the case of a
machine learning to play chess, which will play games against itself in
order to maximize the number of victories. These methods are widely
used for learning to play games or in the domain of robotics. So far, they
have had few applications to brain diseases and will not be covered in
the rest of this book.

3.1.4. Discussion
Unsupervised learning is obviously attractive because it does not re-
quire labels. Indeed, acquiring labels for a training set is usually time-
consuming and expensive because the labels need to be assigned by a
human. This is even more problematic in medicine because the labels
must be provided by experts in the field. It is thus in principle at-
tractive to adopt unsupervised strategies, even for tasks which could be
framed as supervised learning problems. Nevertheless, up to now, the
performances of supervised approaches are often vastly superior in many

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 15

Clustering

Figure 8: Clustering task. The algorithms automatically identifies three

groups (corresponding to the red circles) from unlabelled examples (the
blue dots). The groups are not known a priori. Figure courtesy of Johann
Faouzi.

Machine Learning for Brain Disorders, Chapter 1

16 Colliot

applications. However, in the past years, an alternative strategy called

self-supervised learning, where the machine itself provides its own super-
vision, has emerged. This is a promising approach which has already led
to impressive results in different fields such as natural language process-
ing in particular [32, 33, 34].

3.2 Overview of the learning process

In this section, we aim at formalizing the main concepts underlying most
supervised learning methods. Some of these concepts, with modifications,
also extend to unsupervised cases.
The task that we will consider will be to provide an output, denoted as
y, from an input given to the computer, denoted as x. At this moment,
the nature of x does not matter. It can for example be any possible
photograph as in the example presented in Figure 9. It could be also
be a single number, a series of numbers, a text . . . For now, the nature
of y can also be varied. Typically, in the case of regression, it can be
a number. In the case of classification, it corresponds to a label (for
instance the label “cat” in our example). For now, you do not need to
bother about how these data (images, labels . . . ) are represented in a
computer. For those without a background in computer science, this will
be briefly covered in subsection 3.3.
Learning will aim at finding a function f that can transform x into y,
that is such that y = f (x). For now, f can be of any type, just imagine
it as an operation that can associate a given x with a given y. In chapter
3, the functions f will be artificial neural networks. Learning aims at
finding a function f which will provide the correct output for each given
input. Let us call the loss function and denote ℓ a function that measures
the error that is made by the function f . The loss function takes two
arguments: the true output y and the predicted output f (x). The lower
the loss function value, the closer the predicted output is to the true
output. An example of loss function is the classical least squares loss
ℓ(y, f (x)) = (y − f (x))2 but many others exist. Ideally, the best function
f would be the one that produces the minimal error for any possible
input x and associated output y, not only those which we have at our
disposal, but any other possible new data. Of course, we do not have
any possible data at our disposal. Thus, we are going to use a set of
data called the training set. In supervised learning, this set is labelled,
i.e. for each example in this set, we know the value of both x and y. Let
us denote as (x(1) , y (1) ), . . ., (x(n) , y (n) ) the n examples of the training set
which are n pairs of inputs and outputs. We are now going to search
for the function f that makes the minimum error over the n samples of

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 17

Training set: labeled examples

Cats

Input
Racoons

Learning
Trees

Cat Output

Figure 9: Main concepts underlying supervised learning, here in the

case of classification. The aim is to be able to recognize the content of
a photograph (the input x) which amounts to assigning it a label (the
output y). In other words, we would like to have a function f that
transforms x into y. In order to find the function f , we will make
use of a training set (x(1) , y (1) ), . . ., (x(n) , y (n) ) (which in our case is a
set of photographs which have been labelled). All images come from
https://commons.wikimedia.org/ and have no usage restriction.

Machine Learning for Brain Disorders, Chapter 1

18 Colliot

the training set. In other words, we are looking for the function which
minimizes the average error over the training set. Let us call this average
error the cost function:
n
1X
ℓ y (i) , f (x(i) )


J(f ) =
n i=1

Learning will then aim at finding the function fˆ which minimizes the
cost function:
n
1X
fˆ = arg min ℓ y (i) , f (x(i) )


f ∈F n i=1

In the above equation, arg min indicates that we are interested in the
function f that minimizes the cost J(f ) and not in the value of the cost
itself. F is the space that contains all admissible functions. F can for
instance be the set of linear functions or the set of neural networks with
a given architecture.
The procedure that will aim at finding f that minimizes the cost is
called an optimization procedure. Sometimes, the minimum can be find
analytically (i.e. by directly solving an equation for f ) but this will rarely
be the case. In other cases, one will resort to an iterative procedure (i.e.
an algorithm): the function f is iteratively modified until we find the
function which minimizes the cost. There are cases where we will have
an algorithm that is guaranteed to find the global minimum and others
where one will only find a local minimum.
Minimizing the errors on the training set does not guarantee that
the trained computer will perform well on new examples which were
not part of the training set. A first reason may be that the training
set is too different from the general population (for instance, we have
trained a model on a dataset of young males and we would like to apply
it to patients of any gender and age). Another reason is that, even if
the training set characteristics follow those of the general population,
the learned function f may be too specific to the training set. In other
words, it has learned the training set “by heart” but has not discovered a
more general rule that would work for other examples. This phenomenon
is called overfitting and often arises when the dimensionality of the data
is too high (there are many variables to represent an input), when the
training set is too small or when the function f is too flexible. A way to
prevent overfitting will be to modify the cost function so that it not only
represents the average error across training samples but also constrains
the function f to have some specific properties.

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 19

3.3 Inputs and features

In the previous section, we made no assumption on the nature of the
input x. It could be an image, a number, a text, etc.
The simplest form of input that one can consider is when x is a
single number. Examples include age, clinical scores . . . However, for
most problems, characterization of a patient cannot be done with a single
number but requires a large set of measurements (Table 1). In such a case,
the input can be a series of numbers x1 , . . . , xp which can be arranged
into a vector:  
x1
x =  ... 
 
xp

Table 1: Example where the input is a series of number. Here each

patient is characterized by several variables

Age (years) Height (cm) Weight (kg)

Patient 1 52.5 172 52
Patient 2 75.1 182 78
Patient 3 32.7 161 47
Patient 4 45 190 92

However, there are cases where the input is not a vector of numbers.
This is the case when the input is a medical image, a text or a DNA
sequence for instance. Of course, in a computer, everything is stored as
numbers. An image is an array of values representing the gray-scale in-
tensity of each pixel (Figure 10). A text is a sequence of characters which
are each coded as a number. However, unlike in the example presented
in Table 1, these numbers are not meaningful by themselves. For this
reason, a common approach is to extract features, which will be series
of numbers that meaningfully represent the input. For example, if the
input is a brain magnetic resonance image (MRI), relevant features could
be the volumes of different anatomical regions of the brain (this specific
process is done using a technique called image segmentation which is cov-
ered in another chapter). This would result in a series of numbers that
would form an input vector. The development of efficient methods for
extracting meaningful features from raw data is important in machine
learning. Such an approach is often called feature engineering. Deep
learning methods allow for avoiding extracting features by providing an
end-to-end approach from the raw data to the output. In some areas,
this has made feature engineering less important but there are still ap-

Machine Learning for Brain Disorders, Chapter 1

20 Colliot

plications where so-called “handcrafted features” are competitive with

deep learning methods.
0 0 0 0 0 0 1 2 4 4 6 4 5 3 5 4 3 5 5 5 4 5 4 3 1 0 0 0 0 0 0 0
1 0 0 1 0 0 0 1 1 1 1 1 0 0 0 -1 -2 -1 -1 -1 -1 0 1 1 1 1 0 0 0 1 0 0
1 1 1 1 1 1 1 1 1 -1 -2 0 7 16 20 24 29 30 32 33 25 14 3 -1 0 1 1 1 1 1 1 1
1 1 1 1 1 1 1 -1 0 14 33 50 49 44 42 38 38 35 34 33 37 45 47 27 1 0 1 1 1 1 1 1
1 1 1 0 1 1 -1 11 45 51 39 37 33 33 49 61 57 60 70 61 49 46 33 46 51 9 -1 1 1 1 1 1
1 1 1 1 1 -1 18 46 33 34 53 63 77 74 79 84 74 82 83 85 73 74 68 37 40 58 16 -1 1 1 1 1
1 1 1 1 -1 18 40 37 61 70 87 89 88 87 83 83 87 86 86 88 73 81 78 72 35 33 54 6 0 1 1 1
1 1 1 0 8 39 39 70 87 86 89 90 89 89 79 79 89 84 87 89 85 76 79 66 55 36 44 30 -2 1 1 1
1 1 1 0 38 38 71 84 87 84 81 80 82 84 83 92 92 85 83 84 79 75 84 70 81 56 31 45 4 0 1 1
1 1 0 12 43 54 82 83 84 76 80 89 91 90 93 94 94 91 83 80 82 90 90 86 76 78 46 39 22 -1 1 1
1 1 -2 28 42 64 83 81 75 82 93 91 87 88 84 79 77 78 93 94 83 83 87 92 88 73 74 36 41 2 1 1
1 1 -1 35 46 79 83 83 78 92 90 67 82 87 87 85 85 70 67 90 95 89 92 89 76 70 86 50 37 16 -1 1
1 0 7 52 44 73 87 82 83 88 91 81 87 94 94 91 92 91 81 69 93 96 86 81 75 86 85 63 35 31 -1 1
1 -1 15 63 26 75 88 88 86 85 90 90 85 92 95 93 91 90 89 84 93 89 81 86 90 90 81 83 38 37 1 1
1 0 7 61 20 27 58 77 83 83 81 79 84 84 84 91 92 87 80 88 85 86 86 86 88 90 88 85 46 39 4 1
1 1 0 38 74 21 19 41 43 46 51 67 67 68 66 79 90 73 71 79 78 83 86 89 90 90 87 85 53 36 3 1
2 1 4 17 36 30 27 10 7 10 12 44 106 96 74 71 83 77 82 86 81 78 78 80 87 82 82 83 44 33 1 1
0 4 25 41 40 30 36 13 5 3 5 12 53 106 87 55 85 89 87 94 88 85 82 78 75 74 76 57 39 21 -1 1
13 29 49 57 37 38 35 26 12 14 29 16 20 64 68 48 67 85 86 92 92 91 87 83 80 73 40 26 48 3 1 3
50 50 66 77 50 39 21 25 27 15 33 50 50 57 39 37 39 53 76 87 89 89 88 85 84 56 28 55 27 -1 0 19
26 20 23 46 44 38 40 39 47 50 40 35 72 77 62 54 41 52 83 85 86 86 85 75 52 37 76 57 1 0 1 6
-1 29 58 64 39 37 43 36 40 41 30 19 59 76 68 55 44 58 72 73 72 65 54 52 52 73 76 20 -1 1 1 1
-1 27 73 73 29 47 50 47 72 66 36 45 58 69 65 45 51 61 89 61 60 62 75 71 75 73 58 1 1 2 2 1
-1 27 72 36 26 54 69 75 77 79 79 73 70 70 62 46 48 63 64 57 70 72 69 64 73 77 28 -2 1 1 1 1
0 23 46 46 70 74 71 71 72 73 74 73 66 64 61 42 38 59 65 57 69 67 63 69 68 67 5 1 1 1 1 1
-1 44 68 39 61 62 61 63 65 69 75 73 73 67 62 47 45 46 49 60 69 61 63 67 69 40 -2 1 1 1 1 1
-1 20 82 33 44 64 63 62 61 62 73 73 62 67 62 50 57 53 46 56 60 56 61 65 67 20 -1 1 1 1 1 1
1 5 56 42 20 60 70 64 60 61 64 68 41 50 63 47 43 45 38 53 50 51 61 61 55 8 0 1 1 1 1 1
-1 18 58 39 27 51 67 60 55 60 59 60 35 29 58 48 44 46 40 54 53 49 59 56 42 2 1 1 1 1 1 1
-1 24 57 35 23 46 57 57 56 58 55 55 37 31 57 47 39 39 40 51 52 49 56 53 30 -1 1 1 1 1 1 1
0 8 44 30 30 48 50 49 50 49 43 49 47 26 49 37 33 40 34 45 46 46 47 44 22 0 2 1 1 1 1 1
0 0 5 17 21 24 25 24 27 26 26 23 13 12 27 23 21 24 26 28 28 30 25 25 13 0 1 0 0 0 0 0

Figure 10: In a computer, an image is represented as an array of num-

bers. Each number corresponds to the gray-level of a given pixel. Here
the example is a slice of an anatomical MRI which has been severely un-
dersampled so that the different pixels are clearly visible. Note that an
anatomical MRI is actually a 3D image and would thus be represented by
a 3D array rather than by a 2D array. Image courtesy of Ninon Burgos.

3.4 Illustration in a simple case

We will now illustrate step-by-step the above concepts in a very simple
case: univariate linear regression. Univariate means that the input is a
single number as in the example shown in Figure 7. Linear means that the
model f will be a simple line. The input is a number x and the output is a
number y. The loss will be the least squares loss: ℓ(y, f (x)) = (y−f (x))2 .
The model f will be a linear function of x i.e. f (x) = w1 x + w0 and
corresponds to the equation of a line, w1 being the slope of the line
and w0 the intercept. To further simplify things, we will consider the
case where there is no intercept, i.e. the line passes through the origin.
Different values of w1 correspond to different lines (and thus to different
functions f ) and to different values of the cost function J(f ), which can
be in our case rewritten J(w1 ) since f only depends on the parameter w1
(Figure 11). The best model is the one for which J(w1 ) is minimal.
How can we find w1 such that J(w1 ) is minimal? We are going to use
dJ
the derivative of J: dw 1
. A minimum of J(w1) is necessarily such that
dJ
dw1
= 0 (in our specific case, the inverse is also true). In our case, it is

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 21

! $(#!)
' " ="
3

2 ' " = 0.5"

X
1 ' " = 0.3"
X
X
1 2 3 " 0 1 #!

Figure 11: We illustrate the concepts of supervised learning on a very

simple case: univariate linear regression with no intercept. Training sam-
ples correspond to the black circles. The different models f (x) = w1 x
correspond to the different lines. Each model (and thus each value of the
parameter w1 ) corresponds to a value of the cost J(w1 ). The best model
(the blue line) is the one which minimizes J(w1 ), here it corresponds to
the line with a slope w1 = 1.

dJ
possible to directly solve dw 1
= 0. This will nevertheless not be the case
in general. Very often, it will not be possible to solve this analytically.
We will thus resort to an iterative algorithm. One classical iterative
method is gradient descent. In the general case, f depends not on only
one parameter w1 but on a set of parameters (w1 , . . . , wp ) which can be
assembled into a vector w. Thus, instead of working with the derivative
dJ
dw1
, we will work with the gradient ∇w J. The gradient is a vector that
indicates the direction that one should follow to climb along J. We will
thus follow the opposite of the gradient, hence the name gradient descent.
This process is illustrated in Figure 12, together with the corresponding
algorithm.

4. Conclusion
This chapter provided an introduction to machine learning (ML) for a
non-technical readership (e.g. physicians, neuroscientists . . . ). ML is an
approach to artificial intelligence and thus needs to be put into this larger
context. We introduced the main concepts underlying ML that will be
further expanded in Chapters 2 to 6. The reader can find a summary of
these main concepts, as well as notations, in Box 3.

Machine Learning for Brain Disorders, Chapter 1

22 Colliot

Figure 12: Illustration of the concept of gradient descent in a sim

where the model f is defined using only one parameter w1 . The
$(#updated
w1 is iteratively !) by following the opposite of the gradient

method is gradient descent. In the general case, f depends not

X
one parameter w1 but on a set of parameters (w1 , . . . wp ) which
assembled into a vector w. Thus, instead of working with the de
dJ
dw1
X work with the gradient rw J. The gradient is a vec
, we will
indicates the direction that one should follow to climb along J.
X opposite of the gradient, hence the name gradient
thus follow the
This process is illustrated in Figure 12. The gradient descent al
X
proceeds as follows: # !

repeat
dJ
w1 w1 ⌘ dw 1
until convergence;

Figure 12: Upper panel: where

Illustration
⌘ is of
thethelearning
concept of gradient
rate, descent
i.e. the speed at which w1 will
in a simple case where the model f is defined using only one parameter
dated.
w1 . The value of w1 is iteratively updated by following the opposite of the
gradient. Lower panel: gradient descent algorithm where η is the learning
rate, i.e. the speed at which w1 will be updated.
4. Conclusion
This chapter provided an introduction to machine learning (M
non-technical readership (e.g. physicians, neuroscientists . . . ). M
approach to artificial intelligence and thus needs to be put into th
context. We introduced the main concepts underlying ML that
further expanded in Chapters 2 to 6. The reader can find a sum
these main concepts, as well as notations, in Box 3.

Machine Learning for Brain Disorders, Chapter 1

Machine Learning for Brain Disorders, Chapter 1
Introduction to machine learning 23

Box 3: Summary of main concepts

• The input x

• The output y

• The training samples (x(1) , y (1) ), . . ., (x(n) , y (n) )

• The model: transforms the input into the output

f such that y = f (x)

• The set of possible models F

• The loss: measures the error between the predicted and the
true output, for a given sample
ℓ(y, f (x))

• The cost function: measures the average error across the train-
ing samples
J(f ) = n1 ni=1 ℓ y (i) , f (x(i) )
P

• Learning process: finding the model which minimizes the cost

function
fˆ = arg minf ∈F J(f )

Machine Learning for Brain Disorders, Chapter 1

24 Colliot

Acknowledgments
The author would like to thank Johann Faouzi for his insightful com-
ments. This work was supported by the French government under man-
agement of Agence Nationale de la Recherche as part of the “Investisse-
ments d’avenir” program, reference ANR-19-P3IA-0001 (PRAIRIE 3IA
Institute) and reference ANR-10-IAIHU-06 (Institut Hospitalo-Universitaire
ICM).

References

[1] Samuel AL (1959) Some studies in ma- [11] McCarthy J (1960) Recursive functions of
chine learning using the game of checkers. symbolic expressions and their computation
IBM Journal of research and development by machine, part I. Communications of the
3(3):210–229 ACM 3(4):184–195

[2] Russell S, Norvig P (2002) Artificial intelli- [12] Cardon D, Cointet JP, Mazières A,
gence: a modern approach. Pearson Libbrecht E (2018) Neurons spike
back. Reseaux 5:173–220, URL
https://neurovenge.antonomase.fr/
[3] McCulloch WS, Pitts W (1943) A logical
RevengeNeurons_Reseaux.pdf
calculus of the ideas immanent in nervous
activity. The bulletin of mathematical bio-
physics 5(4):115–133 [13] Rumelhart DE, Hinton GE, Williams RJ
(1986) Learning representations by back-
propagating errors. Nature 323(6088):533–
[4] Wiener N (1948) Cybernetics or Control 536
and Communication in the Animal and the
Machine. MIT press [14] Le Cun Y (1985) Une procédure
d’apprentissage pour réseau à seuil as-
[5] Hebb DO (1949) The organization of behav- symétrique. Cognitiva 85:599–604
ior. Wiley
[15] LeCun Y, Boser B, Denker JS, Henderson
[6] Turing AM (1950) Computing machinery D, Howard RE, Hubbard W, Jackel LD
and intelligence. Mind 59(236):433–360 (1989) Backpropagation applied to hand-
written zip code recognition. Neural com-
putation 1(4):541–551
[7] McCarthy J, Minsky ML, Rochester N,
Shannon CE (1955) A proposal for the
dartmouth summer research project on [16] Matan O, Baird HS, Bromley J, Burges
artificial intelligence. Research Report URL CJC, Denker JS, Jackel LD, Le Cun Y, Ped-
http://raysolomonoff.com/dartmouth/ nault EPD, Satterfield WD, Stenard CE,
boxa/dart564props.pdf et al (1992) Reading handwritten digits:
A zip code recognition system. Computer
25(7):59–63
[8] Newell A, Simon H (1956) The logic theory
machine–a complex information processing [17] Legendre AM (1806) Nouvelles méthodes
system. IRE Transactions on information pour la détermination des orbites des
theory 2(3):61–79 comètes. Firmin Didot

[9] Rosenblatt F (1958) The perceptron: a [18] Pearson K (1901) On lines and planes of
probabilistic model for information storage closest fit to systems of points in space.
and organization in the brain. Psychological The London, Edinburgh, and Dublin philo-
review 65(6):386 sophical magazine and journal of science
2(11):559–572
[10] Buchanan BG, Shortliffe EH (1984) Rule-
based expert systems: the MYCIN exper- [19] Fisher RA (1936) The use of multiple mea-
iments of the Stanford Heuristic Program- surements in taxonomic problems. Annals
ming Project. Addison-Wesley of eugenics 7(2):179–188

Machine Learning for Brain Disorders, Chapter 1

Introduction to machine learning 25

[20] Loh WY (2014) Fifty years of classification [28] Paszke A, Gross S, Massa F, Lerer A,
and regression trees. International Statisti- Bradbury J, Chanan G, Killeen T, Lin
cal Review 82(3):329–348 Z, Gimelshein N, Antiga L, et al (2019)
Pytorch: An imperative style, high-
[21] Quinlan JR (1986) Induction of decision performance deep learning library. In: Ad-
trees. Machine learning 1(1):81–106 vances in neural information processing sys-
tems, vol 32, pp 8026–8037
[22] Vapnik V (1999) The nature of statistical
learning theory. Springer [29] Chollet F, et al (2015) Keras. URL https:
//github.com/fchollet/keras
[23] Boser BE, Guyon IM, Vapnik VN (1992) A
training algorithm for optimal margin clas-
sifiers. In: Proceedings of the fifth annual [30] Shortliffe E (1976) Computer-based medical
workshop on Computational learning the- consultations: MYCIN. Elsevier
ory, pp 144–152
[31] Mitchell T (1997) Machine learning. Mc-
[24] Pedregosa F, Varoquaux G, Gramfort Graw Hill
A, Michel V, Thirion B, et al (2011)
Scikit-learn: Machine learning in python. [32] Mikolov T, Chen K, Corrado G, Dean J
the Journal of machine Learning research (2013) Efficient estimation of word repre-
12:2825–2830 sentations in vector space. arXiv preprint
arXiv:13013781
[25] Krizhevsky A, Sutskever I, Hinton GE
(2012) Imagenet classification with deep
convolutional neural networks. In: Ad- [33] Radford A, Narasimhan K, Sali-
vances in neural information processing sys- mans T, Sutskever I (2018) Improv-
tems, vol 25, pp 1097–1105 ing language understanding by gen-
erative pre-training. URL https:
[26] LeCun Y, Bengio Y, Hinton G (2015) Deep //cdn.openai.com/research-covers/
learning. Nature 521(7553):436–444 language-unsupervised/language_
understanding_paper.pdf
[27] Abadi M, Agarwal A, et al (2015) Ten-
sorFlow: Large-scale machine learning on [34] Devlin J, Chang MW, Lee K, Toutanova
heterogeneous systems. URL https://www. K (2018) Bert: Pre-training of deep bidi-
tensorflow.org/, software available from rectional transformers for language under-
tensorflow.org standing. arXiv preprint arXiv:181004805

Machine Learning for Brain Disorders, Chapter 1