Introduction to Machine Learning (for AI)

Introduction to Machine Learning

Dr. Matias Valdenegro

November 14, 2024

1 / 72
Today’s Lecture

• In this lecture you will learn the basic Machine

Learning concepts.
• We will see the three paradigms of ML: Supervised,
Unsupervised, and Reinforcement Learning.
• We will discover what are features and a bit on how
to engineer them...
• And you will learn basic data preprocessing and
normalization applied to features, together with the
biggest challenges in ML.

2 / 72
Outline

1 Introduction

2 Basic Machine Learning Concepts

3 Feature Engineering

4 Data Preprocessing and Normalization

5 Challenges in Machine Learning

3 / 72
Introduction and Basic ML concepts

• Machine learning is a field that studies algorithms

that can learn from data.
• In this part of the lecture we will cover the basic
concepts in detail.
• In this course we will cover classical ML algorithms,
which are the base for more complex algorithms like
neural networks. The basic concepts are the same,
so covering them in detail is very important.

4 / 72
Introduction and Basic ML concepts

Image from https://7wdata.be/big-data/10-companies-using-machine-learning-in-cool-ways/

5 / 72
What is required to reach True AGI?

Adaptivity or Adaptability
Be able to learn from previous errors and improve overall performance
by learning from experience.

Learning
Learning is the process of acquiring new understanding, knowledge,
behaviors, skills, values, attitudes, and preferences.

Generalization
Lessons and learned knowledge should be general enough that they
also work in (new) previously unseen environments.

6 / 72
The Qualification Problem
Modeling the real world is very difficult. The Qualification
problem is that it is impossible to enumerate or list all
pre-conditions that are required for an action in the real
world to succeed and have the intended effect. For
example:
• To attend an online lecture, you need to have an
internet connection.
• You also need a power source for your computer.
• And the computer should not be broken.
• And the network interface should also work correctly.
• And so on...
Actions can also have unintended consequences, and
enumerating these is also impossible.
7 / 72
Data Driven Paradigms
In our time, the AI field is a combination of numerous
techniques. The (currently) dominant set of techniques is
machine learning, meaning that algorithms learn from
data and/or experience.

Large quantities of data are used to specify a task that

the agent should learn. For some tasks, this data might
require labels provided by a human, while other tasks
might learn without labels.

This alleviates the Qualification problem, as data is better

(but not perfect) in capturing the realities of our world.

8 / 72
Learning Paradigms

Supervised Learning
In supervised learning, the agent learns through supervision, that is, a
teacher tells the agent what is the correct answer given some input
or stimuli, and the agent should learn the mapping between input and
desired output.

Supervised Learning is generally limited by the quantities of data that

are labeled.

Producing labels for datasets is expensive (in time and money), and
in many cases there might not be a correct answer to provide a label.

9 / 72
Learning Paradigms
Unsupervised Learning
In Unsupervised Learning, with contrast of Supervised Learning, the
agent learns from data directly, without any kind of supervision.
There is no correct answer. The agent should learn structure,
properties, and relationships in the data itself.

Unsupervised Learning is not limited by labels, thus it can exploit

large amounts of data that are not labeled, the only limitation is how
to capture additional data of interest.

The problem in Unsupervised Learning is generally algorithmic, it is

difficult to extract patterns and properties of data without
supervision.

The ideal Unsupervised Learning method would make no assumptions

on data.
10 / 72
Learning Paradigms
Unsupervised Learning

Figure: Example of convolutional features that can be learned

using unsupervised learning on images. The algorithm learns
that images consists of edges, then groups edges into parts,
and groups them into more complex objects.
11 / 72
Learning Paradigms
Reinforcement Learning
In Reinforcement Learning, the agent interacts with an environment,
sensing and performing actions that change the environment, and
learns with a reward signal. This signal tells the agent how it is
performing, but does not explicitly say which is the correct answer or
the best action.

The agent learns a policy that maximizes a specified criterion, for

example, to maximize the sum of rewards over a whole interaction
episode.

This paradigm is commonly used for sequential decision making

tasks, for example, controlling a robot to pick and object and place it
somewhere else, or play video games.

12 / 72
Learning Paradigms
Reinforcement Learning
Figure: Diagram
representing how the agent
interacts with the
Environment environment, by performing
an action, and then
receiving a reward and a

Action
Rew new state observation. The
a rd process is repeated
Interpreter sequentially until a final
state is reached, or the
Stat
e episode ends. The agent
learns through multiple
episodes.
Agent
Figure from https://en.wikipedia.org/wiki/Reinforcement˙learning
13 / 72
Learning Paradigms

Transfer Learning
It is a combination of previous paradigms, where learned knowledge is
used to learn or improve on new tasks. Generally this means part of
the learned knowledge transfers from previous tasks into new ones.

This is possible because tasks usually have some common knowledge

that can be exploited. For example, in an image understanding
problem, if one learns a image classifier for dogs/cats, the features
that are learned in this case are also useful for classifying birds, as
images naturally have features like edges, parts, etc.

14 / 72
Learning Paradigms

Combinations of Previous Paradigms

Self-Supervised
The agent/algorithm supervises itself by means of
solving a pretext task. This task can be automatically
computed and solved, but generally it needs to be
designed by humans.
Semi-Supervised Learning
The agent/algorithm receives multiple kinds of
supervision at the same time, for example it could be
doing unsupervised learning on a set of images, while
using that knowledge and performing supervised
learning on a different set of images.

15 / 72
Learning Paradigms

Combinations of Previous Paradigms

Weak Supervision
The agent receives supervision but it is weak, meaning
that the kind of supervision received is not enough to
completely solve the task.
For example for object detection, bounding box labels
are needed (box coordinates and class information),
and in weakly supervised object detection, only class
labels are supervised.

16 / 72
The Cake/Lecake - Relative Importance of
Learning Paradigms

Figure: The Lecake, A common representation on the value or

relative importance of each learning paradigm, made by Yann
LeCun.
17 / 72
What is not Machine Learning?

Image from https://xkcd.com/1838/

18 / 72
What is not Machine Learning?

• Machine Learning is not magic, data is not the new

oil/electricity, and proper methods still need to be
developed.
• It is a rigorous field that is more than just applying
methods/libraries without thinking.
• If methods or evaluation are not performed correctly,
unrealistic performance can be obtained. Then you
only fool yourself.
• Many startups expect Machine Learning to magically
solve their problems. But it does not work that way.

19 / 72
Core Idea of Machine Learning
Machine Learning, as a field, has a bunch of methods,
techniques, methodologies, but in the end the core idea
or most important concept is:

Generalization
This means you train a model on a dataset, and it should
work (make correct predictions) on very different input
data, be useful outside of its training space.

We will define generalization more formally in coming

lectures.

20 / 72
Outline

1 Introduction

2 Basic Machine Learning Concepts

3 Feature Engineering

4 Data Preprocessing and Normalization

5 Challenges in Machine Learning

21 / 72
Basic ML concepts
Model
A model is the actual equations that define how the input is
transformed into output predictions. For example:
X
f (x) = wi xi + b
i

We denote models as f (x) where x is the input, and models with

probability as p(y |x). Note that ML models are usually predictive
ones.

Parameters or Weights
These are the variables that change during the training process,
implicitly encoding the knowledge that is learned. We usually denote
these with letters w and b (weights and biases).

22 / 72
Basic ML concepts

Loss Function
A function that makes a comparison between the model’s prediction
and the ground truth labels. It provides direct supervision to the
model so it can learn successfully. Loss functions are selected to learn
a particular task, in coming lectures we will cover the full details.

Training Data
The data that you use to train the model, represented
mathematically as the x and y variables, where x represents inputs,
and y are the labels associated with that input.

23 / 72
Basic ML Concepts
Classification
Classification is the task where the output variable and/or labels are
discrete, that is, the opposite of continuous. A model that performs
classification is called a classifier.

The set of possible outputs for the classifier is called the class set,
and the individual elements are called classes.
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

1
Figure: Binary Classification Figure: Multi-Class
(Two classes) Classification ( > 2 classes)

24 / 72
Basic ML Concepts

Regression
Regression is the task where the output variable and/or labels are
continuous. A model that performs regression is called a regressor.

1
0.8 Training Data
0.6 Model Prediction
0.4
0.2
0
0
0.2
0.4
0.6
0.8
1

Figure: Regression in 2D

25 / 72
Basic ML Concepts

Multi-Task Learning
• It is also possible to perform/learn more than one task at a
time.
• For example, performing object detection requires object
recognition (classification) and bounding box estimation
(regression).
• The idea is that tasks have some input or features in common,
so performance is higher when learning these tasks together,
than each individual task in isolation.

26 / 72
Learning as Optimization
All machine learning models learn from data, using a
general formulation that consists of the following
optimization problem:
n
X
∗
θ = min L(f (xi ), yi , θ) (1)
θ
i
The solution to the optimization problem is θ∗ , the
weights/parameters that produce the best predictions,
according to what the loss function scores between the
prediction f (xi ) and the labels yi .

The loss function is usually designed to be used to learn a

specific task. A common source of errors is choosing the
wrong loss function.
27 / 72
Why does this work?
• This might look as simple function approximation (it
is), but when the model is a complex non-linear
function, and the training data is a large dataset,
they can contain and represent complex relationships
that are not easy for humans.
• The performance of a model increases with the
number of data points in the training set. We will go
deeper in this direction in future lectures.
• In any case this requires a considerable amount of
engineering work, including capturing and labeling
data, selecting preprocessing pipelines and models,
and training.

28 / 72
Trainable and Non-Trainable Parameters
In the previous formulation, we usually train the
parameters using an optimization algorithm, but there
could be multiple types of parameters.

Trainable Parameters
Parameters that are directly trained using an optimization algorithm.

Non-Trainable Parameters
Parameters that are trained but not using an optimization algorithm,
for example, models that indirectly learn parameters from data that
passed through them.

In transfer learning, parameters can also be frozen, where they are

kept constant, and new models are trained on top of learned
features. This is done so knowledge is preserved.
29 / 72
Train and Test Datasets
Datasets are usually used for multiple purposes, like
training and evaluation. This requires that once you have
a set of data points, it needs to be split into at least two
datasets:
Training Set
The set that will be used for training using the optimization process,
it usually is split from around 70% to 80% of the data available.

Test Set
The set that will be used for evaluation of the trained model, it
should be a independent dataset containing samples that have not
been seen during training. It can be the remaining 20% to 30% of
the available data.

30 / 72
Linear Separability
This is important as the basic classification concept uses
separating hyperplanes.
Definition
Given points xi ∈ Rn in a high dimensional space (features), with
binary labels yi ∈ [0, 1] they are linearly separable if there exists a
weight vector w ∈ Rn and real number k such as:

w · xi > k ∀xi ∈ [xi | yi = 0]

w · xi < k ∀xi ∈ [xi | yi = 1]

In simple words, there is a hyperplane that separates both classes

y = 0 and y = 1 in the feature space.

Note that the separating hyperplane w does not have to

be unique. This concept can be extended for > 2 classes.
31 / 72
Linear Separability

Blue points are class 0,

and red points are class
1.

Figure from https://en.wikipedia.org/wiki/Linear˙separability

32 / 72
Not Linearly Separable
1

0.8

0.6

0.4

0.2

0
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1
Figure: Example of Not linearly separable data for Binary
Classification.

The role of Feature Engineering for Classification is to

transform features to be linearly separable.
33 / 72
Outline

1 Introduction

2 Basic Machine Learning Concepts

3 Feature Engineering

4 Data Preprocessing and Normalization

5 Challenges in Machine Learning

34 / 72
Basic ML Concepts
Features
This is a very important and highly misunderstood term. A feature is
a measurement, measurable property, or characteristic of a
phenomenon or physical object.

The inputs to a machine learning model are feature values, that are
used to perform classification or regression. The quality of the
features is what defines the performance of your model.
• Good features will lead to good performance of your model.
• Bad features will prevent your model from making correct
predictions. Model might not train at all.
• Features should be de-correlated and independent for best
results. If features are related to each other, then they are
redundant and not useful for learning a model.

35 / 72
Basic ML Concepts
Features - Examples
Images
Pixel values, edges, parts, color(s), distribution of pixel
values (histograms), distribution of edges, gradient of
image.
Text
Frequency of words, language, presence or absence of
specific words, # of grammatical errors, overall
Sound
Frequency response (Fourier transform coefficients),
response to filter banks, signal to noise ratio, length of
sound sequences.

Note that features should be representable as numerical data.

36 / 72
Basic ML Concepts
Feature Vectors
To finalize the concept of a feature, we need to talk about the
concept of a feature vector. This is a vector with elements
corresponding to individual features. For example let’s try to predict
the rent for an apartment, our features could be:
A. Area in m2 (continuous).
B. Location in the city (categorical).
C. Distance to nearest train station (continuous).
D. Build year (categorical).

Then a feature vector could be x = [A, B, C ], or x = [A, C , D], or

x = [A, D], etc.

Designing features and feature vectors is called feature engineering.

37 / 72
Types of Features
Continuous
Real numbers. For example, distances, floating point numbers,
convolutional feature maps, gradients, etc.

Discrete
Integer numbers that can be ordered and are countable. For example,
number of shoes in your apartment, most currencies (€, $), number
of persons living in a house, population of a country, etc.

Categorical
Categories, where no order can be established. For example, colors,
nationality, course code, etc.

38 / 72
Obtaining Features

There is a whole field of Feature Engineering, to obtain

high quality features for tasks. Specially in NLP and
Computer Vision.

The quality of your features related to a specific task,

defines the performance of your machine learning model.
Better features means better performance, and viceversa.

Features can also be (non-linear) transformations of raw

data (pixels, sound, text, etc). This is what Deep
Learning does.

39 / 72
Feature Transformations
A simple way to ”improve” your features is to make
aggregation or non-linear transformations of features. For
example:
Aggregation
Mean, standard deviation, variance, skew,
kurtosis, etc.
Polynomials
For a scalar feature x, use x, x 2 , x 3 , x 3 , ..., x p .
The polynomial degree p is a
hyper-parameter.
Frequency
Fourier transform of your data is commonly
used as features, with a frequency cutoff.
40 / 72
Not Useful Features

• Features that are linearly correlated with each other.

• Features that do not change over the whole dataset
(constant).
• Features with many missing or unclear values.
• Features that do not predict the target value of
interest.
• Random features.

41 / 72
Outline

1 Introduction

2 Basic Machine Learning Concepts

3 Feature Engineering

4 Data Preprocessing and Normalization

5 Challenges in Machine Learning

42 / 72
Pre-Processing
In general data needs to be pre-processed and normalized
before being used with any model. This is because when
learning a model, it will only work with data that is similar
to the training data distribution, so normalizing allows
data to be more similar to this distribution, and not to
vary too much.

• Pre-processing can also discard invalid data, select

which features will be used, and put data in a
common format.
• Train and test splits are also part of pre-processing.
• Depending on the task (like regression), outputs can
also be post-processed to produce ideal predictions.
43 / 72
Normalization
Normalization is the process where input features are
standarized to a common range.
• This is required as features might be at different
scales (numerical range), and the model might assign
importance to each feature according to the highest
value.
• In general training a model with normalized data is
much easier, so it is a recommended first step.
• Target values can also be normalized, for example, to
make regression easier (by limiting the model output
range).
• Intuition is that each (normalized) feature should
have similar range.
44 / 72
Normalization
• We cover a set of normalization techniques. A very
important point is that normalization generally
captures some statistics from the training set, so any
normalization should be applied in the training set,
and then in the test set (using train statistics) and in
any new data when making predictions.
• Normalization allows all features to have similar
ranges, which makes training smoother. Models
trained on unnormalized data usually perform worse.
• Sometimes it is difficult to normalize when the
training and test sets vary considerably.

45 / 72
Normalization

Min-Max
For each feature x, compute the following:
x − min x
xnorm = (2)
max x − min x
This transforms all values to range inside [0, 1]. The minimum
feature value is mapped to zero, and the maximum feature is mapped
to one. Intermediate values are mapped to (0, 1).

There is also a transformation to the [a, b] range:

x − min x
xnorm = (b − a) +a (3)
max x − min x

46 / 72
Normalization

Standard or Mean-Std
For each feature x, compute the following:

x − mean(x) x − µx
z= = (4)
std(x) σx

This transforms all values to an approximate range, usually [−3, 3],

depending on the distribution of x. You might recognize that this is
the same as the z-score used in a normal/gaussian distribution.
The mean of z is 0.0, and the standard deviation and variance is 1.0.

47 / 72
Normalization

Unit Vector
For a feature vector x, compute the following:
x xi
v= ={ }i (5)
||x|| ||x||

pP Where ||x||
This makes the data lie on a n dimensional hyper-sphere.
is a vector norm, for example the L2 norm ||x|| = 2
i xi

48 / 72
Normalization - Whitening

Whitening is a more complex transformation of

multi-dimensional data X , with the purpose of
de-correlating and setting the covariance matrix Σ of the
data to the identity matrix.

For zero-centered (mean substracted) data X , the basic

transformation is:

Z = WX (6)
The matrix W is called a whitening matrix, and should
follow that W T W = Σ−1 .

49 / 72
Normalization - Whitening

Depending on the choice of W , there are different types

of whitening:
ZCA Whitening
1
Select W = Σ− 2
PCA Whitening
Compute W using Principal Component
Analysis (PCA).
Note that ZCA and PCA whitening produce different
results but in both cases the data is normalized to identity
covariance.

50 / 72
Normalization - PCA Whitening
Here the matrix W is computed using the eigenvalue
decomposition on Σ, which is a matrix decomposition in
the form:
Σ = UΛU T (7)
Where U contains the eigenvectors of Σ as columns, and
Λ is a diagonal matrix containing the eigenvalues of Σ.
This can be computed using functions like
np.linalg.eig. Then W is computed as:
1
W = Λ− 2 U T (8)

This sets the covariance of X to be a identity matrix.

51 / 72
Normalization - Images

For specialized data types, there are specific

normalization methods. For images we have:
Histogram Equalization This adjust contrast in the image
so the distribution (histogram) of gray levels
is uniform. This makes sure that all gray
levels are equally represented in the
normalized image. It is useful for when the
image has dominant shadows or highlights.

52 / 72
Normalization - Images

Noise Removal Noise can be removed using a median or

gaussian filter, with a careful design to not
remove important image features.
Resizing Images in a dataset can vary in size, and they
are usually normalized to a standard size.
Common values are 224 × 224 or 256 × 256.
Normalized size must be carefully chosen to
not lose image features.

53 / 72
Normalization - Histogram Equalization

0 50 100 150 200 250

0 50 100 150 200 250

Figure: Original image and Figure: Equalized image and

histogram (red). The black histogram (red). The black
curve is the cumulative curve is the cumulative
distribution. distribution. 54 / 72
Normalization - Audio

Resampling The sampling rate of the audio signal can

also be normalized, common values at 44.1
kHz or 22.05 kHz.
Audio Length The length (in time) of audio signals can
also be normalized, in a similar way to image
size normalization.

55 / 72
Normalization - Audio
Amplitude Normalization The overall volume of audio can
also be normalized, to make sure all samples
in the dataset have similar volume. This is
done in two ways:
1. Peak Normalization. Normalizing to
the maximum value across time samples.
This is similar to min-max scaling.
2. Loudness Normalization. Normalizing
to a aggregate loudness of the whole
audio signal across time. For example it
can be the root mean square (RMS), or
a value related to human perception of
loudness.
56 / 72
Normalization - Text
Text is the most difficult to normalize, since many factors
can vary:
• Stop words are usually removed.
• Numbers, dates, acronyms and abbreviations can be
standarized to common terms (like transforming 10
to Ten, dates to ISO formats, and acronyms
removed or expanded.
• Non alphanumeric characters are usually removed.
• Specific words can be standardized, specially if they
have multiple meanings, depending on the
application.
• Spelling across variations of the same language can
also be normalized, for example British vs Australian
vs American English.
57 / 72
Normalization - Categorical Variables

Categorical variables are usually not numeric, for example,

when a variable can take a value from a pre-defined set of
values. A common mistake is to just assign numerical
values to each of the categorical values, but this biases
the model, as it thinks higher value means more
importance.

So you cannot use categorical variables by converting

them to integers. This is a common mistake.

58 / 72
Normalization - Categorical Variables
An unbiased way to encode categorical variables is to use
one-hot encoding.

If the variable has N possible values and we encode

category i , then a vector of N elements is used to
represent it, where all elements are 0.0, except for the
i -th value which is set to 1.0.

An example:
• Cat = [1.0, 0.0, 0.0]
• Dog = [0.0, 1.0, 0.0]
• Bunny = [0.0, 0.0, 1.0]

59 / 72
Normalization - Regression

One important detail, is that output values (labels) can

also be normalized, particularly in the case of regression.
• If your labels have a limited range, normalizing can
teach the model what values are allowed. Models can
also be configured to output in a normalized range.
• One should also make sure to remove or clip output
values that are outside of the expected ranges.
• Correct normalization of the output and labels can
the process of learning a model much easier.

60 / 72
Normalization - Which to use?

We have covered a lot of normalization methods.

Choosing them is difficult, usually we experiment during
ML model development and choose what performs best.
• In many cases it is trial and error.
• Experience tells us what to try first.
• Exploratory analysis of your data can reveal
properties that should be normalized. Always look at
your data before trying to learn models from it.

61 / 72
Normalization - Importance

What can happen if you do not normalize your data?

• Model underperforms.
• Training fails (loss does not decrease and metrics do
not improve).
• Model makes prediction that do not make sense,
specially if.
• Much higher chance of getting stuck in local minima.
The loss surface/landscape and the random weight
initialization are both affected by normalization of the
input/output.

62 / 72
Normalization - (Negative) Example

Figure: No Normalization Figure: X and Y are normalized.

Notice the difference in fitting a non-linear model without
(left) and with (right) normalization applied. It is literally
night and day. The model shown here has uncertainty but
the principle is the same.

63 / 72
Outline

1 Introduction

2 Basic Machine Learning Concepts

3 Feature Engineering

4 Data Preprocessing and Normalization

5 Challenges in Machine Learning

64 / 72
Challenges in Machine Learning

Quality of Data
The biggest bottleneck is obtaining high quality data (including
features and labels) for the task of interest. Without data, a model
cannot be learned.

Low quality data (incorrect labels, unclear or useless features, missing

values, etc) hurts the performance of your model.

Well known saying, Garbage in, Garbage out.

65 / 72
Challenges in Machine Learning

Model Misspecification
Model misspecification is using the wrong model for a given dataset.

In simple words, you have some features/data, a task and a target to

predict. We do not know what would be the best model for this task.

In general you do not know the mathematical model that generated

the data. Specially for highly dimensional data.

So we usually do trial and error and research to find out this.

66 / 72
Challenges in Machine Learning

Generalization
Overall I think the biggest issue currently in AI is the lack of
generalization, that means, solutions are given to specific problems,
but general AI solutions (true AGI) does not exist yet.
• Alpha Go and variations (Alpha Go Zero and Alpha Zero) only
solved playing the game of Go, but these advances have not
benefited other applications.
• Deep Blue can play chess, but it has not served to solve some
of humanities biggest issues (like peace, hunger, wealth
distribution, etc).
• All models overfit to some degree, but some are useful.

67 / 72
Challenges in Machine Learning
Explainability
Most AI methods overall are Black Boxes, meaning it is difficult to
interpret their decisions, or to get an explanation on how a decision is
made.
• Explanations and interpretability are required for AI to gain
human trust, and it is even legally required in some countries
(like in the EU with the GDPR).
• Explaining black box method is very difficult, as computational
concepts do not always map to human understandable
concepts. The use of non-linear methods also makes it difficult.
• Building methods that are easily interpretable (AKA white box
methods) is also very difficult, and in many cases they work but
are less accurate than black box models.

68 / 72
Challenges in Machine Learning
Bias, Safety, and Trustworthiness
As AI is used for real world applications, specially ones involving
humans, it is very important to ensure that it will not perform
unexpected actions or behave abnormally. Right now there are AI
systems that are actively hurting people, for example:
• In the US, poorly understood AI systems are being used to
decide if a person will commit a crime, or how long their
sentence should be.
• AI systems being used for monitoring of populations, violating
their human rights.
• AI systems used by banks to review and recommend on
decisions for loans, and by companies to decide who to
interview for a job.
• Discrimination by DUO and the Toeslagenaffaire (Benefits
scandal). 69 / 72
Questions to Think About

1. What is a feature?
2. Describe linear separability in your own words.
3. What is the need for normalization and
pre-processing of features?
4. Can regression labels/targets be normalized?
5. What factors can hurt the performance of a Machine
Learning model?
6. Your model does not learn, what is the most likely
cause?

70 / 72
Take Home Messages

• There are three paradigms in Machine Learning,

Supervised, Unsupervised, and Reinforcement.
• Generalization is the biggest issue/core idea of
Machine Learning.
• Two main tasks: Classification and Regression.
• Learning is done through an optimization problem.
• Normalization and pre-processing of features and
labels is very important.
• Machine Learning is not magic, there are many
challenges.

71 / 72
Questions?

72 / 72