Introducción al

aprendizaje automático
TC3002B
Basic Concepts
Introduction

2023FJ - [email protected] 2
 – Ability to
– use percepts from the outside world
– not only for reacting,
What is – but for improving actions in future events.

learning?
– Implies that we know when and how to use this new knowledge.
– When: pattern detected
– How: algorithm created.

2023FJ - [email protected] 3
 – Example:
– Imagine a supermarket chain with a hundred of stores selling
groceries to millions of customers.
– Each sale has a lot of data that can be analised and converted
into information.
What is – These information can be used to give people suggestions
when buying.
machine
learning? – If we knew who would buy an item, we would just write
code for the computer to remind them.

– Because we do not know, we collect data and hope to

extract enough information to recommend articles to
people.

2023FJ - [email protected] 4
 – Example:
– In RoboCup agents play soccer.
– There are 11 players against 11 players.
– Each team has its own strategy for playing soccer.

What is – If we knew which strategy a team is using, we would play a

counter-attack strategy to stop them.
machine
learning? – Because we do not know their strategy, we collect data and try to
extract enough information to detect their strategies.

– Once strategies are detected and classified, we could select the

best strategy to exploit this knowledge.

2023FJ - [email protected] 5
 – The computer algorithm should be able to:
– Identify patterns in the data (When)

– Construct a good and useful approximation of the solution to the

What is ML? … problem (How)

2023FJ - [email protected] 6
 – “Machine learning uses data and answers to discover rules behind
a problem” Chollet (2017)

– “Machine learning is programming computers to optimize a

performance criterion using example data or past experience.”
Alpaydin, E. (2004)

What is ML? … – Has a model defined for some parameters.

– Learning is the execution of a computer program to optimize the
parameters of the model using training data or past experience.

– Two types of models:

– Predictive model: predictions in the future.
– Descriptive model: gain knowledge from data.

2023FJ - [email protected] 7
 – “A computer program is said to learn from experience E with
respect to some class of tasks T and performance measure P, if its
performance at tasks in T, as measured by P, improves with
experience E.” Mitchell, T. (1997)

– Example: handwriting recognition

What is ML? … – Task T: recognizing and classifying handwritten words within
images.
– Performance measure P: percent of words correctly classified
– Training experience E: a database of handwritten words with given
classifications

2023FJ - [email protected] 8
Learning Agent

2023FJ - [email protected] 9
 1. Which components of the performance element should be
learned?
Design of a
learning 2. What feedback is available to learn these components?
element
3. What representation is used for the components?

2023FJ - [email protected] 10
 – What can be learned?
– Direct mapping from conditions on the current state to actions.

– Means to infer relevant properties of the world from the percept

sequence.
Components
of the – Information about the way world evolves and the results of possible
actions agent can take.
performance
– Utility information indicating the desirability of world states.
element
– Action-value information indicating desirability of actions.

– Goals that describe classes of states whose achievement maximizes

the agent’s utility.

2023FJ - [email protected] 11
 – Components can be learned from appropriate feedback.
– Example: training Tae Kwon Do, Driving a Taxi.

– Type of feedback:
– The most important factor in determining the nature of the learning
Feedback problem.

– Three cases:
1. Supervised learning
2. Unsupervised learning
3. Reinforcement learning

2023FJ - [email protected] 12
 – Learning a function from examples of its inputs and outputs.
– There is an input X, an output Y, and the task is to learn mapping
from input to output.

– Outputs values can be provided

– By a supervisor – someone feed the output.
Supervised – By the environment – detected by sensors.

Learning
– Examples:
– Learn a condition-action rule for punching.
– Learn to differentiate between a dog and a cat.
– Regression
– Classification.

2023FJ - [email protected] 13
 – Learning patterns in the input when no specific output values are
supplied.

– Aim: to find regularities in the input.

Unsupervised
learning – Example:
– Learn to separate colors.
– Learn when it might rain.
– Learn how to detect people that will not pay their credit cards.

2023FJ - [email protected] 14
 – Mix between supervised and unsupervised learning

– Some data is labelled – usually a very small part

Semi-
supervised – Labelled data is used to create more data
learning
– Learner learns to:
– Generate labelled data and to
– Detect regularities in the input

2023FJ - [email protected] 15
 – The output of the system is a sequence of actions.

– Uses rewards to guide the sequence of actions

– These actions are part of a policy.

– A single action is not important.
Reinforcement – The policy is what must be learned.

learning – Agent must learn from reinforcement which actions are best, i.e. the
policy.

– Examples:
– Playing chess.
– Driving politely.
– Robot navigation.

2023FJ - [email protected] 16
 – Polynomials

– Propositional logic

Representation – Predicate calculus

of the learned
information – Bayesian networks

– Neural networks

– Etc.

2023FJ - [email protected] 17
 – Learning associations
– Learn how people associate
elements (ex. buying – Knowledge extraction
groceries) – Learning a rule from data – it
explains the data
– Rules are a form of data
– Classification compression
– Learn to classify elements in
Applications of different categories
– Outlier detection
machine – Prediction
– Data that does not belong to
a class
learning – Learn to predict if some
action will happen
– Regression problems
– Learn the curve that best fits
– Pattern recognition a function to a set of points
– Learn to find familiar
patterns (characters, faces,
objects, etc.)

2023FJ - [email protected] 18
 – Artificial Intelligence

– Bayesian methods

– Computational complexity theory

– Control theory
ML is
multidisciplinary – Information theory

– Philosophy

– Psychology and neurobiology

– Statistics

2023FJ - [email protected] 19
Designing a learning
system
Introduction

2023FJ - [email protected] 20
 Hyper-
Data Model parameter
Collection Fitting tuning

ML Process
Data Model
Preparation Evaluation

2023FJ - [email protected] 21
 1. Choosing the training experience
1. Feedback
2. Control of sequence of examples
3. Distribution of examples

Designing a 2.
1.
Choosing the target function
Function that is operational
learning
system 3. Choosing a representation for the target function
1. Expressive representation

4. Choosing a function approximation algorithm

1. Estimating training values
2. Adjusting the weights

2023FJ - [email protected] 22
How to build a
dataset

2023FJ - [email protected] 23
Structured Data
ML models learn from examples
Each example is called an instance or
pattern
Dataset is formed with multiple
examples
Structured Data is organized in rows
and columns
A column is called a feature

Images, videos and text are called

Unstructured Data
https://machinelearningmastery.com/wp-content/uploads/2013/12/Table-of-Data-Showing-an-Instance-Feature-and-Train-Test-Datasets.png

2023FJ - [email protected] 24
Dataset
organization
and division
Training set is usually 80% of
original set

Test set is usually 20%

Validation set is usually 20% of

training set

https://miro.medium.com/max/585/0*lbveKaL-MGRgppD8.png

2023FJ - [email protected] 25
 – When data is too big and we
can’t pass all data to computer
at once.

– One Epoch is when an entire

dataset is passed forward and
backward through the learning
model only once.

Epoch, batch – Batch size: divide dataset into

number of batches or sets or
& iteration parts.

– Iterations is the number of

batches needed to complete
one epoch.

– The number of batches is equal

to number of iterations for one
epoch.

https://towardsdatascience.com/epoch-vs-iterations-vs-batch-size-
4dfb9c7ce9c9
2023FJ - [email protected] 26
 – Data Wrangling
– Data might be in different
files
– Cleaning, structuring,
enriching raw data – Data Preparation
– Assure quality and useful – Analysis and optimization of
data features
– Select/remove features
– Consider prediction needs
– Data Cleansing and computation time
Data – Missing values (delete?)
– Unwanted characters
Preprocessing – Unwanted elements

https://miro.medium.com/max/666/0*ScsuON73dMJDC9XO.png

2023FJ - [email protected] 27
Complete data
science
pipeline

https://developer.ibm.com/articles/ba-intro-data-science-1/

2023FJ - [email protected] 28
Model
Evaluation

2023FJ - [email protected] 29
 – N x N matrix Actual values
– N = number of classes

Predicted values
Confusion – Evaluates performance of a
Matrix classification model

– Compares actual target https://www.python-course.eu/images/confusion.matrix_image.png

values with predicted values

2023FJ - [email protected] 30
 – True Positive (TP)
– Predicted value matches
actual
– Both were positive
– True Negative (TN)
– Predicted value matches
actual
– Both are negative
Binary
confusion – False Positive (FP)
– Type I error
matrix – Predicted value falsely
predicted
– Actual value Negative
– False Negative (FN)
– Type II error
https://cdn.analyticsvidhya.com/wp-content/uploads/2020/04/Basic-Confusion-matrix.png – Predicted value falsely
predicted
– Actual value Positive

2023FJ - [email protected] 31
 – Accuracy
– Fraction of predictions model correctly classified

!"##$%& '#$()%&)"*+
– 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = ,"&-. */01$# '#$()%&)"*+
Confusion
matrix metrics – For binary classification
!"#!$
– 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = !"#!$#%"#%$

– Very simple, but does not take into consideration class imbalances
and data unevenly distributed

2023FJ - [email protected] 32
 – Precision
– Proportion predicted positives identified correctly
!"
– 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
!"#$"

– Recall (Sensitivity)
– Proportion actual positives identified correctly
!"
– 𝑅𝑒𝑐𝑎𝑙𝑙 = !"#$%

Confusion – Specificity
– Proportion actual negatives identified correctly
matrix – 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = !%#$"
!%

metrics…
– Precision used when FP is a higher concern than FN
– From the predicted positives, how many are really positive?

– Recall used when there is a high cost associated with FN

– How many positive were correctly classified?
– A higher recall ensures more actual positive values are being identified

2023FJ - [email protected] 33
 – F1 Score
– Helps understand balance between Precision and Recall

2 '#$%)+)"* × #$%-.. ,6
– 𝐹1 = =2𝑥 =
Confusion !
3
!
"#$%&& '"#$()(*+
'#$%)+)"*5#$%-.. !
,6 5 ,(86589)

matrix – Values range from 0 to 1

metrics… – A value close to 1 means it is a better model

– Used when
– there is a need to balance this two metrics
– Not easy to decide if Type I or Type II errors is preferred

2023FJ - [email protected] 34
 – When a classifier is not reporting the values we desire, we can
move the threshold for classification

– Moving threshold can increase/decrease recall, precision and

ROC & AUC specificity values

– ROC and AUC can help us determine the best threshold

– Receiver Operator Characteristic (ROC)
– Area Under the Curve (AUC)

2023FJ - [email protected] 35
ROC
– Summarizes all confusion matrices produced with
different thresholds

– Diagonal line is where TP rate is equal to FP rate

– Points above the diagonal represent a good classifier

– The best classifier would be (1,0)

– The value of the threshold is the value used in the

classifier that produced the ROC graph

2023FJ - [email protected] 36
 AUC

– Helps to compare different ROC

graphs

– The greater the value of the AUC, the

better the model is for classifing that
data

2023FJ - [email protected] 37
Issues in
machine
learning

2023FJ - [email protected] 38
 – What algorithms exist for learning general target functions from
specific training examples?

– How much training data is sufficient?

Issues in – When and how can prior knowledge guide the process of generalizing
from examples?
machine
learning – What is the best strategy for choosing a useful training experience?

– What is the best way to reduce the learning task to one or more
function approximation problems?

– Can the learner learn to represent the target function?

2023FJ - [email protected] 39
 – Alpaydin, Ethem (2004). Introduction to Machine Learning. The MIT Press.

– Mitchell, Tom (1997). Machine Learning. WCB McGraw-Hill.

– Edwards, Gavin (2018). Machine Learning, an introduction. Towards Data

References Science (https://towardsdatascience.com/machine-learning-an-
introduction-23b84d51e6d0)

– Josh Starmer (2019). ROC and AUC clearly explained! StatQuest with Josh
Starmer, YouTube Channel

2023FJ - [email protected] 40