2024-03-29T01:32:32Zhttp://oai-repositori.upf.edu/oai/requestoai:repositori.upf.edu:10230/331072018-01-24T08:35:45Zcom_10230_20650com_10230_16441col_10230_33097
2017-10-27T10:19:58Z
urn:hdl:10230/33107
Resolution of concurrent planning problems using classical planning
Furelos Blanco, Daniel
Tutor: Anders Jonsson
Treball fi de màster de: Master in Intelligent Interactive Systems
In this work, we present new approaches for solving multiagent planning and temporal
planning problems. These planning forms are two types of concurrent planning,
where actions occur in parallel. The methods we propose rely on a compilation to
classical planning problems that can be solved using an off-the-shelf classical planner.
Then, the solutions can be converted back into multiagent or temporal solutions.
Our compilation for multiagent planning is able to generate concurrent actions that
satisfy a set of concurrency constraints. Furthermore, it avoids the exponential
blowup associated with concurrent actions, a problem that many multiagent planners
are facing nowadays. Incorporating similar ideas in temporal planning enables
us to generate temporal plans with simultaneous events, which most state-of-the-art
temporal planners cannot do.
In experiments, we compare our approaches to other approaches. We show that the
methods using transformations to classical planning are able to get better results
than state-of-the-art approaches for complex problems. In contrast, we also highlight
some of the drawbacks that this kind of methods have for both multiagent and
temporal planning.
We also illustrate how these methods can be applied to real world domains like the
smart mobility domain. In this domain, a group of vehicles and passengers must
self-adapt in order to reach their target positions. The adaptation process consists
in running a concurrent planning algorithm. The behavior of the approach is then
evaluated.
2017-10-27T10:19:58Z
2017-10-27T10:19:58Z
2017-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/33107
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/331082018-01-24T08:36:04Zcom_10230_20650com_10230_16441col_10230_33097
2017-10-27T10:28:41Z
urn:hdl:10230/33108
3D pose estimation for symmetric and nonsymmetric objects
Corona Puyane, Enric
Treball fi de màster de: Master in Intelligent Interactive Systems
Supervisor: Sanja Fidler; Co-Supervisor: Coloma Ballester
Autonomous systems have to understand the 3D space, being able to detect objects
and infer their pose to pick them and reliably perform a certain goal-oriented action.
An increasing number of works focus on this topic motivated by self-driving cars
or the Amazon picking challenge. In particular, we focus on pose estimation, a
well-known problem in computer vision and robotics which is essential for object
manipulation. This requires reliable identification of object poses to know how to
pick them up in order to interact with them for a certain goal. Additionally, pose
estimation involves several difficulties. Objects may have rotational symmetries or
their appearance can vary significantly depending on the lighting or occlusions.
A common approach to pose estimation is to first estimate a coarse pose to initialize
ICP and get a fine pose estimation. We follow this idea in this work by
comparing an object in an RGB-D setting to a set of views of the same CAD model
obtained offline. Using Convolutional Neural Networks, we embed the images to a
common space where they can be efficiently compared. Additionally, we propose to
consider symmetries directly in the comparison to avoid inconsistencies in the pose
estimation.
Given the lack of benchmarks with symmetric objects for pose estimation, we obtain
6669 CAD models of very different kinds and generate realistic simulations of tabletop
scenarios to train and test our approach. We also leverage a non-published
dataset of real objects with symmetries. Finally, we infer rotational symmetries in
new CAD models, obtaining a high recall and promising results that suggest further
research.
2017-10-27T10:28:41Z
2017-10-27T10:28:41Z
2017-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/33108
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/331092018-01-24T08:36:06Zcom_10230_20650com_10230_16441col_10230_33097
2017-10-27T10:32:43Z
urn:hdl:10230/33109
Cross-Entropy method for Kullback-Leibler control in multi-agent systems
Cabrero Daniel, Beatriz
Supervisor: Dr. Vicenç Gómez Cerdà; Co-Supervisor: Dr. Mario Ceresa
Treball fi de màster de: Master in Intelligent Interactive Systems
We consider the problem of computing optimal control policies in large-scale multiagent
systems, for which the standard approach via the Bellman equation is intractable.
Our formulation is based on the Kullback-Leibler control framework, also
known as Linearly-Solvable Markov Decision Problems. In this setting, adaptive
importance sampling methods have been derived that, when combined with function
approximation, can be effective for high-dimensional systems. Our approach
iteratively learns an importance sampler from which the optimal control can be
extracted and requires to simulate and reweight agents’ trajectories in the world
multiple times. We illustrate our approach through a modified version of the popular
stag-hunt game; in this scenario, there is a multiplicity of optimal policies
depending on the “temperature” parameter of the environment. The system is built
inside Pandora, a multi-agent-based modeling framework and toolbox for parallelization,
freeing us from dealing with memory management when running multiple
simulations. By using function approximation and assuming some particular factorization
of the system dynamics, we are able to scale-up our method to problems
with M = 12 agents moving in two-dimensional grids of size N = 21×21, improving
on existing methods that perform approximate inference on a temporal probabilistic
graphical model.
2017-10-27T10:32:43Z
2017-10-27T10:32:43Z
2017-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/33109
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/331102018-01-24T08:36:14Zcom_10230_20650com_10230_16441col_10230_33097
2017-10-27T10:36:48Z
urn:hdl:10230/33110
Computational comparative analysis of the Twitter networks of the 2015 and 2016
Spanish national elections
Gallego Gamo, Helena
Supervisors: Andreas Kaltenbrunner, Pablo Aragón, Vicenç Gómez
Treball fi de màster de: Master on Intelligent and Interactive Systems
In the last years, Spanish politics have transitioned from bipartidism to multipartidism.
This change led to an unstable situation which finally evolved to the rare
scenario of two general elections in the period of six months. The two elections
had a main difference: the two biggest left-wing parties formed a coalition in the
second election while they had run separately in the first one. In the second election
and after merging, the coalition lost around one million votes contradicting opinion
polls. In this study, community analysis in the retweet networks of the two online
campaigns is performed in order to assess whether activity in Twitter reflects the
outcome or parts of the outcomes of both elections. The results show that the leftwing
parties lost more online supporters than the other parties. Furthermore, an
inspection of the Twitter activity of the supporters unveils a decrease in engagement
especially marked for the smaller party in the coalition, in line with post-electoral
traditional polls. The clusters obtained with the community detection method are
also used to situate in the ideological spectrum a set of Spanish media sources and to
understand their audiences and behavioral differences when replying or retweeting
them.
2017-10-27T10:36:48Z
2017-10-27T10:36:48Z
2017-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/33110
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/331112018-01-24T08:36:13Zcom_10230_20650com_10230_16441col_10230_33097
2017-10-27T10:52:19Z
urn:hdl:10230/33111
A generative model of user activity in the Integrated Learning Design Environment
Bas Serrano, Joan
Treball fi de màster de: Master on Intelligent and Interactive Systems
Supervisors: Vicenç Gómez and Davinia Hernández-Leo
The objective of this project has been to build a generative model of the user activity
in the Integrative Learning Design Environment (ILDE) able to describe the data of
different communities and which can be used for both to gain understanding about
the data and to test hypothetical situations.
The model that we present is called Hierarchical Multivariate Hawkes Model and
works with a two layer procedure that first draws the beginning of working sessions
and then fills these sessions with events of different kinds using a Multivariate
Hawkes Model.
In the project we first make an statistical temporal analysis of the data, to understand
it and see the important features to be modeled, then we introduce the model,
validate it, and show some of its applications.
Through these steps it has been shown that the model is able to reproduce satisfactorily
the sequences of events produced by the ILDE users and it can be easily used
to tackle real problems that would be difficult to face with typical statistical tools.
2017-10-27T10:52:19Z
2017-10-27T10:52:19Z
2017-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/33111
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/331122018-01-24T08:36:02Zcom_10230_20650com_10230_16441col_10230_33097
2017-10-27T10:57:45Z
urn:hdl:10230/33112
Towards suicide prevention: early detection of depression on social media
Leiva Aranda, Víctor
Supervisor: Ana Maria Freire Veiga
Treball fi de màster de: Master in Intelligent Interactive Systems
The statistics presented by the World Health Organization inform that 90% of the suicides
can be attributed to mental illnesses in high-income countries. Besides, previous
studies concluded that people with mental illnesses tend to reveal their mental condition
on social media, as a way of relief. Among all these users of social media platforms,
adolescents are the most frequent ones. Hence, these previous studies drive us through
the detection of depression on social media as a first step against online suicidal behaviour.
Thus, the main objective of this work is the analysis of the messages that
user posts online, sequentially through a time period, and detect as soon as possible
if the user is at risk of depression. Our preliminary experiments report the impact of
sentiment analysis techniques and a combination of machine learning algorithms for
detecting users with depression in Reddit.
2017-10-27T10:57:45Z
2017-10-27T10:57:45Z
2017-10-27
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/33112
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/332942018-01-24T08:36:59Zcom_10230_20650com_10230_16441col_10230_33097
2017-11-21T10:57:10Z
urn:hdl:10230/33294
Term extraction and document similarity in an Integrated Learning Design Environment
Martínez Rodríguez, Alberto
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Davinia Hernández-Leo i Horacio Saggion
The Integrated Learning Design Environment is a social platform focused in supporting
teachers in the computer-assisted design of Learning activities. In this platform, teachers
and course designers can contextualize, author and share their designs within their
community. This social component, of the ILDE, would benefit from the application of
Information Retrieval and Natural Language Processing techniques to facilitate teachers
and course designers to find shared designs as fast and efficient as possible. In this work,
we use Natural Language Processing to classify learning designs written in Catalan, get
the content of the users, parse this content with Freeling and extract education domainspecific
terminology from the documents. To extract the terminology, a combination of
two methods is used. The first method uses the Multilingual Central Repository ontology
to check if a term belongs to any of four pedagogical fields. The second methodology,
computes the tf-idf of all the documents terms using a non-domain-specific corpus, the
Catalan Wikipedia. This work also discusses the potential of the proposed combination
of methods to retrieve simple and complex terms from documents. The resulting
combined method distributes the weight of each method in the extraction process to assign
a score to each retrieved term. After this process of extracting education domain-specific
terminology from different ILDE documents, it has been created a Document Similarity
Application addressed to teachers and course designers. This application allows users to
search documents based on the similarity between these documents and another document
of the same ILDE community. Besides, given a document, users can visualize the
education terminology that belongs to that document. Finally, users can also search for
certain documents using a terminology-based query to obtain a set of documents and their
similarity with respect to that query.
2017-11-21T10:57:10Z
2017-11-21T10:57:10Z
2017-11-21
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/33294
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/356952018-11-06T10:31:55Zcom_10230_20650com_10230_16441col_10230_33097
2018-11-05T12:35:29Z
urn:hdl:10230/35695
Early detection of eating disorders in reddit
Mayans Yern, Marc
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Ana Freire, Diana Ramírez-Cifuentes
This thesis proposes an approach for the early detection of Anorexia Nervosa (AN)
on social media. Our method is based on machine learning techniques using the
processed texts written by social media users. This method relies on a set of features
based on domain-specific vocabulary, topic modelling, psychological processes and
linguistic information extracted from the users’ writings. Moreover, other features
are studied in order to exploit all the dataset resources. Additionally, we have
compared some of the most known learning algorithms and we have introduced a
minimum amount of information threshold to avoid some false positive predictions.
This approach penalises the delay in the detection of positive cases in order to classify
the users at risk as early as possible. By the early identification of anorexia, along
with an appropriate treatment, the speed of recovery and the likelihood of staying
free of the illness improves. The results of this thesis showed that our proposal is
suitable for the early detection of AN symptoms in social media. Further research
in this topic is needed to solve the problems stated in this project.
2018-11-05T12:35:29Z
2018-11-05T12:35:29Z
2018-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/35695
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/356962018-11-06T10:31:29Zcom_10230_20650com_10230_16441col_10230_33097
2018-11-05T12:40:21Z
urn:hdl:10230/35696
Information extraction from user-generated content in the classical music domain
Porcaro, Lorenzo
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Horacio Saggion
The applications of Information Extraction (IE) on User-generated Content (UGC) have
widely benefited from the emergence of microblogging services in the last decade. In
particular, Twitter has been at the center of attention of many studies because of its
widespread use and easy accessibility. Among the several fields which have benefited from
this source, in particular Named Entity Recognition (NER) has demonstrated how
challenging can be obtaining useful information from the noisy space of tweets . From
another perspective, recently in the field of Music Information Retrieval (MIR) researches
have shown how NLP techniques such as IE and NER can be an important resource to
improve accuracy and precision in tasks like Music Recommendation, Artist Similarity or
Genre Classification . The objective of this thesis is to investigate methods to extract
information from user-generated content in a specific channel related to Classical Music,
BBC Radio 3 , through the use of NER techniques. We investigate how state-of-the-art
methods in NER can be applied to detect entities in the music domain, and how contextual
information can contribute with NER in this particular case.
2018-11-05T12:40:21Z
2018-11-05T12:40:21Z
2018-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/35696
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/356972018-11-06T10:30:51Zcom_10230_20650com_10230_16441col_10230_33097
2018-11-05T12:43:53Z
urn:hdl:10230/35697
Combining radiomics and disease state index for interactive patient space visualization
Martín Isla, Carlos
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Karim Lekadir
With the increasing number of data routinely acquired in clinical practice, a great variety
of predictive models for automated medical diagnosis based on quantitative feature
extraction are being implemented.
Domain analysts should be able to interact and explore through different views to make
further discovery of, and insights into, the quantitative data. They can obtain a better
understanding of the data and their structures and contribute their domain expertise to the
knowledge discovery process.
For this research, we will take advantage of the use of radiomics for the acquisition of
large amounts of relevant data from cardiac images that typically fail to be appreciated by
the naked eye. The combination with Disease State Index, a predictive model that involves
not only a diagnosis but a sense of progression in the disease, leads to a 3D space where
the patients are explored and compared, as well as their individual features.
This configuration will result in a clinical tool that allows the user to explore
radiomics features in a serie of interactive panels and take supervised decisions, but also is
useful for automatic diagnosis and patient stratification. Last but not least, it draws
conclusions about the relevance of different radiomics classes, segmentations, and cardiac
phases for automatic heart disease classification.
2018-11-05T12:43:53Z
2018-11-05T12:43:53Z
2018-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/35697
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/356982018-11-06T10:30:25Zcom_10230_20650com_10230_16441col_10230_33097
2018-11-05T12:51:32Z
urn:hdl:10230/35698
Feature selection from large-scale radiomics data: Application to heart disease diagnosis
Izquierdo Morcillo, Cristian
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Karim Lekadir, Irem Cetin
Radiomics have become in the past years one of the most interesting fields to be studied and analyzed in medicine. Since their first implementations, researchers have been trying to design new algorithms to develop this tool in order to predict with higher accuracy the appearance of some diseases like cancer. In this thesis we will focus in a method for feature selection, one of the most challenging questions that come up when working with radiomics. The aim of this new algorithm is to improve the accuracy and efficiency by making use of parallel computing. We will face, as well, one of the main issues of the new century, dealing with the Big Data.
2018-11-05T12:51:32Z
2018-11-05T12:51:32Z
2018-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/35698
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/356992018-11-06T10:29:47Zcom_10230_20650com_10230_16441col_10230_33097
2018-11-05T12:55:22Z
urn:hdl:10230/35699
Automatic concept extraction from biomedical material
Marin, Albert
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Leo Wanner, Jérôme Noailly
Natural Language Processing is a vibrant field of computer science that provides
computers with the ability of understanding human language. In the field of medical
data, there is a demanding need to lower the amount of documents clinicians and
researchers need to manage in order to learn new concepts to improve their day-today
practice. The research presented in this thesis aims at the design and evaluation
of an algorithm based on neural networks that will extract the relevant entities from
biomedical papers in order to reduce the amount of time needed for reading papers.
Of all the topics in medicine that can take advantage of this thesis, the one it has
been chosen in particular is the one of intervertebral discs. One of the reasons
is the availability of experts on the topic in the current university. Moreover, it
is a very interesting field as cells that form part of this structure have different
properties based on their location. This makes it indeed a complex task to retrieve
the relevant information because depending on the considered region some properties
will be prominent whereas in other they might not be that relevant.
The methodology used in the process it has been to use some off-the-shelf libraries
already implemented in Java as a baseline and then use python to code a new architecture
modifications to allow the algorithm to detect the relevant named entities.
The results are compared with the gold standard obtained from the experts in the
field and the conclusions are drawn from the observations.
2018-11-05T12:55:22Z
2018-11-05T12:55:22Z
2018-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/35699
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/361022018-12-18T02:31:48Zcom_10230_20650com_10230_16441col_10230_33097
2018-12-17T09:38:41Z
urn:hdl:10230/36102
Cooperation is the rule, not the exception: reinforcement learning in the Battle of the Sexes
Puig Camps, Bernat
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Vicenç Gómez Cerdà i Martí Sanchez Fibla
Society is highly influenced by conventions, which are a form of cooperation. In many
situations, individuals act together for the benefit of the group. This phenomenon
is easy to understand when all individuals share the same interest. However, when
there exists conflict, it is not clear if altruism is required or pure self-interest can
lead to cooperation. The repeated version of the Battle of the Sexes game can
summarize this situation. Although conflict is present, players need to cooperate to
obtain good rewards. Here we show experimentally that two selfish reinforcement
learning agents learn to cooperate in this conflictive scenario. We found that two
Q-learning agents playing this game modeled as a Markov Game reach a cooperative
fair solution. That is, two agents that learn based solely on their own self-interest
end up cooperating. Furthermore, we found that Q-learning is able to converge in
this multi-agent situation. Our results demonstrate that cooperation among individuals
in this particular conflictive scenario can be explained by means of pure
self-interest. Moreover, cooperation in this setting is the rule, not the exception as
the convergence to it is robust to parameter asymmetry between agents. We also
introduced opponent modeling into the players as a Beta binomial model. It worked
well in modeling the adversary but agents fail to properly exploit that knowledge.
2018-12-17T09:38:41Z
2018-12-17T09:38:41Z
2018-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/36102
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/361032018-12-18T02:31:52Zcom_10230_20650com_10230_16441col_10230_33097
2018-12-17T09:47:07Z
urn:hdl:10230/36103
Policy gradient methods on a multi-agent game with Kullback-Leibler costs
Requena Pozo, Borja
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Vicenç Gómez i Martí Sanchez Fibla
In nature we find all kinds of multi-agent systems sustained upon cooperative behaviours.
In this work, we study multi-agent systems by means of the Stag-Hunt
game, which presents a conflict between mutual benefit and personal risk. In particular,
we consider the probabilistic inference approach for reinforcement learning
on a grid-based variant of this game. We analyze the behavior of two different
policy gradient algorithms in the presence of function approximation: the standard
REINFORCE algorithm and the Cross-Entropy (CE) method, which differ on the
functional form of the loss. However, even though both REINFORCE and CE share
the same global optimal solution, we have found that REINFORCE behaves too
greedily compared with CE. In agreement with previous results based on probabilistic
graphical models, we obtain two different qualitative optimal solutions (riskand
payoff-dominant) as a function of a temperature parameter, whose transition
is better observed using the CE method. We also analyze the difference between
using or not path-cost, in addition to the end-cost. It is known that adding pathcost
makes the problem harder using an explicit probabilistic graphical model, since
it increases its tree-width. Nevertheless, we observe the opposite effect for policy
gradient methods, for which path-cost enhances the performance of the resulting
controls in all circumstances. This is explained because the samples used by policy
gradients are generally more informed with path-cost. Finally, we also consider a
distributed version of the algorithm, with partial observability and feature sharing
between the agents. In this setting, we show the feasibility of generalizing to larger
grids using training data from smaller grids.
2018-12-17T09:47:07Z
2018-12-17T09:47:07Z
2018-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/36103
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/425382019-12-19T16:42:03Zcom_10230_20650com_10230_16441col_10230_33097
2019-10-29T10:31:25Z
urn:hdl:10230/42538
Predicting user satisfaction to optimize AP selection in WLANs using Random Forests
Carrascosa Zamacois, Marc
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Boris Bellalta
Nowadays, it is common to nd WiFi networks that have a central controller connected
to all Access Points in the network to both organize them and collect relevant
information from them. This creates huge amounts of data which can open new avenues
for Machine Learning to be used in wireless networks, as the amount of data
can be impossible to parse by a human. In this work, we propose a Supervised
Learning model based on Random Forests that can parse all this data and allow
us to predict the satisfaction of all users in the network. To study its performance,
we create a simulated environment from which we can extract a data set to train
the model. Afterwards, we use this model to analyze the importance of the metrics
available and test it in the simulator to con rm its e ectiveness. We then use
the same model to create a process in the simulated central controller that can reassociate
users to Access Points that will o er a better service, reaching a higher
network performance and increasing average user satisfaction.
Avui dia és típic que una xarxa WiFi tingui una controladora central connectada a
tots els punts d’accés de la xarxa, tant per configurar-los com per recollir informació
rellevant de la seva activitat. Aquests processos creen grans quantitats de dades
que ofereixen noves possibilitats per la utilització de Machine Learning en xarxes
sense fils, ja que la quantitat d’informaci ́o generada pot ser impossible de processar
per un humà. En aquest document proposem un model de Supervised Learning
basat en Random Forests que ens permetrà predir la satisfacció de tots els usuaris
en una xarxa. Hem creat una plataforma de simulació de la qual extraiem un
data set amb el qual realitzar l’estudi. Un cop tenim el model, l’utilitzem per a
analitzar les mètriques més importants d’una xarxa i el testem en la simulació per
a confirmar la seva efectivitat. Finalment, utilitzem aquest model per a crear un
procés a la controladora central que reassoci ̈ı usuaris a punts d’accés que puguin
oferir un millor servei, obtenint un major rendiment de la xarxa i incrementant la
satisfacció mitjana per usuari.
Hoy día es habitual que una red WiFi tenga una controladora central conectada
a todos los puntos de acceso de la red, tanto como para configurarlos como para
recoger información relevante de su actividad. Estos procesos crean grandes cantidades de información que ofrecen nuevas posibilidades de utilizar Machine Learning
en redes inalámbricas, ya que la cantidad de información generada puede ser imposible de procesar por un humano. En este documento proponemos un modelo de
Supervised Learning basado en Random Forests que nos permitirá predecir el nivel
de satisfacción de todos los usuarios de la red. Hemos creado una plataforma de
simulación de la cual extraemos un data set con el cual realizaremos el estudio. Una
vez tenemos el modelo, lo usamos para analizar las métricas más importantes de
una red y lo testeamos en la simulación para confirmar su efectividad. Finalmente,
utilizamos este modelo para crear un proceso en la controladora central que reasocie
a usuarios a puntos de acceso que puedan ofrecer un mejor servicio, obteniendo un
mayor rendimiento en la red e incrementando la satisfacci ́on media por usuario.
2019-10-29T10:31:25Z
2019-10-29T10:31:25Z
2019
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/42538
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/425392019-12-19T16:43:19Zcom_10230_20650com_10230_16441col_10230_33097
2019-10-29T10:42:28Z
urn:hdl:10230/42539
Testing optimality in the morphospace of language networks with empirical data
Casal Santiago, Miguel Ángel
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Luís F. Seoane, Ricard Solé
There is an open issue about optimality in human language, which might be behind
some universal features observed across tongues. These features may stem from a
tension between hearers and speakers when trying to minimize costs associated to
their usage of language. Optimality issues might be also critical to understand the
evolution of the language faculty. A toy model introduced by Ferrer i Cancho and
Solé captures the tension between hearers and speakers. In it, tongues are reduced
to a mapping from signals to objects of an external world. Theoretical studies
grounded in information theory followed this study, but the framework remains of
limited empirical use due to the difficulty of building word-objects mappings for real
tongues. There was a recent attempt by Seoane using WordNet, but this database
has some relevant limitations such as the lack of data for some grammatical classes.
In this project, we look at alternative ways to map empirical data from human
languages into the aforementioned least effort information-theory framework. Human
language consistently falls within one of two related categories: i) fairly optimal
(both for hearers and speakers simultaneously) mappings; and ii) less simultaneously
optimal word-object mappings, yet presenting interesting features such as diverse
clustering of concepts and good fitness to Zipf’s law of word frequency.
Our novel empirical analysis of linguistic data allows us to consider more grammatical
classes and to bring together words from different classes coherently. Our
results offer intuitive representations of human languages into an abstract space
where they can be compared with other communication systems. This also offers a
way to quantify the relevance of both conflicting views about optimality in human
language introduced above. As far as optimality could be disregarded, our results
also suggest alternative pressures that might have shaped human language. Future
work will be aimed at scaling the proposed methodology to larger sets of data to
support our findings.
2019-10-29T10:42:28Z
2019-10-29T10:42:28Z
2019
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/42539
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/425402019-12-19T16:44:16Zcom_10230_20650com_10230_16441col_10230_33097
2019-10-29T10:46:08Z
urn:hdl:10230/42540
Musculoskeletal abnormality detection on x-ray using transfer learning
Abreu Dias, Domingo de
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Gema Piella Fenoy, Amelia Jiménez Sanchéz
Radiographic studies are a common technique employed to detect a variety of diseases,
in which the detection of musculoskeletal abnormalities has proven to be a
crucial task. This thesis proposes the use of deep learning techniques to detect
musculoskeletal abnormalities in the MURA dataset, one of the largest collections
of radiographic studies. In particular, we use transfer learning techniques such as
feature extraction and fine-tuning to well-known models for visual tasks such as
InceptionV3, VGG and SqueezeNet, among others. Additionally, we present a tool
based on class activation mappings to aid in visualizing the decision of our models.
The results obtained show that transfer learning techniques can be applied to deep
convolutional neural networks pre-trained on non-medical images, while achieving
comparable results to the state-of-the-art.
2019-10-29T10:46:08Z
2019-10-29T10:46:08Z
2019
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/42540
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/425422019-12-19T16:45:15Zcom_10230_20650com_10230_16441col_10230_33097
2019-10-29T10:53:24Z
urn:hdl:10230/42542
Optimal control using sparse-matrix belief propagation
Iribarne, Albert
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Vicenç Gómez Cerdà
The optimal control framework is a mathematical formulation by means of which
many decision making problems can be represented and solved by finding optimal
policies or controls. We consider the class of optimal control problems that can be
formulated as a probabilistic inference on a graphical model, known as Kullback-
Leibler (KL) control problems. In particular, we look at the recent progress on
exploiting parallelisation facilitated by the graphics processing units (GPU) to solve
such inference tasks, considering the recently introduced sparse-matrix belief propagation
framework [1]. The sparse-matrix belief propagation algorithm was reported
to deliver significant improvements in performance with respect to traditional loopy
belief propagation, when tested on grid Markov random fields.
We develop our approach in the context of the KL-stag hunt game, a multi-agent,
grid-like game which shows two different behavior regimes [2]. We first describe how
to transform the original problem into a pairwise Markov random field, amenable to
inference using sparse-matrix belief propagation and, second, we perform an experimental
evaluation. Our results show that the use of GPUs can bring notable performance
improvements to the optimal control computations in the class of KL control
problems. However, our results also suggest that the improvements of sparse-matrix
belief propagation may be limited by the concrete form of the Markov random field
factors, specially on models with high sparsity within a factor, and variables with
high cardinality.
2019-10-29T10:53:24Z
2019-10-29T10:53:24Z
2019
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/42542
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/425472019-12-19T16:45:59Zcom_10230_20650com_10230_16441col_10230_33097
2019-10-29T11:01:42Z
urn:hdl:10230/42547
A computational and visualisation tool for investigating associations between cardiac
radiomics, risk factors and clinical data
Phloyngam, Naphatthara
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Karim Lekadir, Carlos Martín Isla
Identifying the correlations between radiomics and additional medical, health and
lifestyle factors such as sex, age, BMI, etc. may help in discovering the significant
hidden patterns of data and realizing the causes of the diseases. Also, knowing the
risks in advance is a useful piece of supplementary information which may be used
in addition to medical intervention resulting in preventative measures to reduce the
level of risk or to control prescribed treatments.
In the radiomics and the clinical outcomes datasets, it is hard to identify their correlations
due to the complexity of data, computationally expensive and high number
of possible combination among the choices. Therefore, data pre-processing to keep
only the potential features and data cleaning to deal with missing or non-informative
values are mandatory steps to perform. In addition, applying the powerful Machine
Learning algorithms help to bring the results that even non-specialists in the field
could discover and understand.
This thesis facilitates the discovery of these correlations through the design and
development of an intuitive and interactive web-based tool which dynamically displays
the radiomic feature set alongside the additional medical, health and lifestyle
factors feature set based on the contents of radiomics and clinical data files. The
tool also provides a visualization of the correlation results in an easy to interpret
and meaningful way allowing for effective exploration of any correlations in addition
to cardiovascular risk score calculation.
2019-10-29T11:01:42Z
2019-10-29T11:01:42Z
2019
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/42547
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/425482019-12-19T16:46:41Zcom_10230_20650com_10230_16441col_10230_33097
2019-10-29T11:06:22Z
urn:hdl:10230/42548
Speed traffic sign detection on the CARLA simulator using YOLO
Sánchez Juanola, Martí
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Federico M. Sukno
Research on the self-driving and on the autonomous vehicles field is constantly increasing.
The administrations are putting all their efforts to raise awareness among
drivers about the responsibility that driving a vehicle implies. By their side, researchers
are working hard to make vehicles as safe as possible in order to decrease
the number of fatal accidents in our roads. Around 30% of these accidents have a
key factor: speeding. Since some decades ago until our days, speeding has become a
widely investigated topic by the large automotive companies. The European Union
(EU) has recently announced new measures that will be taken in 2022 aiming to
reduce car accidents. The most relevant one, is the measure that forces new vehicles
manufactured at that year to have an intelligent speed assistance device integrated
to make the driver aware when exceeding the speed limit of the road. Unfortunately,
the oldest cars will remain without that system.
The aim of this project is to contribute to this research by integrating into a realistic
driving simulator a system able to detect and recognise the speed traffic signs of
the road, taking decisions that help the user make the driving process easier and
safer. The simulator used in this study is the CAR Learning to Act (CARLA), an
open-source simulator for autonomous research that mainly consists of two modules:
the CARLA Simulator and the CARLA Python API module. To detect its road
signs, a cutting-edge object-detection algorithm is used: the You Only Look Once
(Yolo) algorithm. Instead of using a sliding window over several locations within an
image, Yolo sees the entire image during the training and testing phases encoding
contextual information about the object classes as well as their appearances. This
characteristic allows it to be extremely fast while an image is being evaluated.
Identifying the speed traffic signs of a road can have a wide range of utilities. In
this project, two applications are exposed: a warning application, to notify the user
that the vehicle speed has overtaken the maximum allowed speed of that road, and
a control application, to reduce the vehicle speed if it exceeds the road speed limit.
Results show that the detection procedure is achieved satisfactorily with a precision
metric value of 0.92. Moreover, the system is tested both on the CPU and on the GPU, making it reproducible in most of the environments. Running it on the
CPU takes a total time of 130 ms, while running it on the GPU, 8 ms are needed
to evaluate the current CARLA scene and determine whether it exists or not any
speed traffic sign.
2019-10-29T11:06:22Z
2019-10-29T11:06:22Z
2019
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/42548
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/425492019-12-19T16:47:16Zcom_10230_20650com_10230_16441col_10230_33097
2019-10-29T11:10:08Z
urn:hdl:10230/42549
Learning state representations and Markov models in football analytics
Soares Afonso, Marielby Mercedes
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Vicenç Gómez Cerdà, Javier Fernández
The increasing availability of spatio-temporal data of football matches in recent years
has prompted the interest of many clubs in performing automated tactical analysis
using machine learning techniques to gain competitive advantage. The low-scoring
nature of the sport, the highly dynamic interactions and the presence of contextual
circumstances that change continuously present challenges for automated analysis.
Using data from football matches of FC Barcelona B, this work aims to automatically
learn a meaningful state representation using high-level features that include
contextual information about the game and to estimate basic Markov models from
the transition probabilities between the states to help coaches to understand player
and team behavior. Multiple clustering techniques have been tested to define states
and a basic Markov model has been estimated for different teams. This allows modeling
how possessions can unfold in any given number of passes, as well as estimating
the probabilities of keeping possession or for it resulting in either turnover, shot or
goal. It has been shown that even a simple model yields useful results for the club
analytics team, that can be used to analyze how a team plays. Also, that this highlevel
representation can help significantly to facilitate the communication between
coaches and analysts thanks to its interpretability.
2019-10-29T11:10:08Z
2019-10-29T11:10:08Z
2019
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/42549
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/463252021-02-05T02:31:08Zcom_10230_20650com_10230_16441col_10230_33097
2021-02-04T07:22:49Z
urn:hdl:10230/46325
Light transport as an optimal control problem
Zarza Pujol, Mariano
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Vicenç Gómez, Gergely Neu
Recently, reinforcement learning (RL) has attracted some attention in the research
field of physically-based rendering. This thesis explores a deeper connection between
both fields. In particular, we identify the light transport equations (LTE) used in
physically-based rendering with the Bellman expectation equations used in optimal
control. With this identification, we formulate an optimal control problem that
can be used for scene rendering based on the framework of linearly-solvable Markov
decision processes. This formulation provides a novel perspective on the use of RL
techniques for rendering.
2021-02-04T07:22:49Z
2021-02-04T07:22:49Z
2021
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/46325
eng
https://creativecommons.org/licenses/by-nc-nd/4.0/
info:eu-repo/semantics/openAccess
Reconeixement-NoComercial-SenseObraDerivada 4.0 Internacional (CC BY-NC-ND 4.0)
oai:repositori.upf.edu:10230/463262021-02-05T02:31:00Zcom_10230_20650com_10230_16441col_10230_33097
2021-02-04T07:29:11Z
urn:hdl:10230/46326
Adaptation of Opeanai Gym for Moveo platform
Acera Mateos, Mario
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Mario Ceresa
In this project we aim to adap Openai Reinforcement Learning platform to the Moveo robot, a low-cost, 3D printed robotic manipulator. This research will widen the field of application of Moveo, concretely for robotics surgeries and also will make the research in robotics available and affordable for a large portion of the research community. Using only open-source technologies we experiment with the different required modules and pose the global architecture to train robots focused on manipulation with Reinforcement Learning techniques. This project is constructed by the combination of three main components: A real robot, a simulation platform and the Reinforcement Learning module. These components are interconnected using the robotic middleware ROS which will provide the corresponding communication interface. As a realistic simulation platform the Gazebo software is used. The Reinforcement Learning module is the OpenAI Gym which interfaces with ROS based robots through the package Openai-ROS. In this document we provide a detailed analysis of the functioning of the package Openai-ROS besides a robotic model in Gazebo with the corresponding functionalities for it to be used along with Openai-ROS and perform the training.
2021-02-04T07:29:11Z
2021-02-04T07:29:11Z
2020
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/46326
eng
https://creativecommons.org/licenses/by-nc-nd/3.0/es
info:eu-repo/semantics/openAccess
Reconeixement-NoComercial-SenseObraDerivada 3.0 Espanya (CC BY-NC-ND 3.0 ES)
oai:repositori.upf.edu:10230/463302021-02-05T02:30:56Zcom_10230_20650com_10230_16441col_10230_33097
2021-02-04T08:39:25Z
urn:hdl:10230/46330
Neural architecture search for detection of deepfakes
Moreno Claver, Jordi
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Vicenç Gómez, Ferran Diego Andilla, Carlos Segura Perales
DeepFakes has become a social problem by slandering the public image of people.
The current and specialized DeepFake detection methods use Hand-Designed Deep
Neural Networks as backbone, which requires a lot of designing effort. We propose
to use a Neural Architecture Search algorithm for DeepFake detection, which aims to
overcome the drawbacks of the Hand-Designed approaches, and has not been proved
yet on DeepFake detection. Additionally, in this work is analyzed the robustness of
the Architecture Search process. We yield comparable results to DeepFake detection
state-of-the-art, that are highly specialized, by adapting Progressive Differential
Architecture Search to DeepFake detection for the first time, using novel techniques
such as data augmentation, multi-label classification and an architecture search
process in the particular domain to improve in terms of performance. Finally, We
have found that the data requirements to obtain a stable architecture are not very
high.
2021-02-04T08:39:25Z
2021-02-04T08:39:25Z
2020
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/46330
eng
https://creativecommons.org/licenses/by-nc-nd/3.0/es
info:eu-repo/semantics/openAccess
Reconeixement-NoComercial-SenseObraDerivada 3.0 Espanya (CC BY-NC-ND 3.0 ES)
oai:repositori.upf.edu:10230/463382021-02-05T02:31:04Zcom_10230_20650com_10230_16441col_10230_33097
2021-02-04T13:59:06Z
urn:hdl:10230/46338
Learning to detect Deepfakes: benchmarks and algorithms
Infante Molina, A. Guillermo
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Vicenç Gómez, Ferran Diego, Carlos Segura
The capabilities of deep-learning tools have led to the emergence of the so-called Deepfakes.
These are a type of videos involving a person whose face has been artificially
forged in one way or another. These videos poses a serious threat to information veracity
and integrity in social media. Therefore, it makes sense that companies and institutions
have a tool available to identify such type of resources in order to take them down from
the Internet. As generation methods have become more and more sophisticated, building
models for the detection of these videos is an increasingly popular area of research.
The task is not easy and requires bringing together several modules as well as taking into
consideration distinct factors.
In this work, we present a survey of the state-of-the-art of current generation and detection
methods. Simultaneously, we analyse the results obtained with different models by formulating
the problem as a binary classification task at a frame level. These results allow
the comparison of some Convolutional Neural Networks architectures as well as several
data augmentation policies. To do so, we have run our models in two different benchmark
datasets: one that is originally from the academia and the another one derived from the
industry. Nonetheless, despite the effort put by researchers on detection methods, more
work has to be done in order to achieve feasible solutions. For example, so far end-toend
trainable models have not yet been accomplished and there exists a generalization
problem in detection models.
2021-02-04T13:59:06Z
2021-02-04T13:59:06Z
2020-11-13
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/46338
eng
https://creativecommons.org/licenses/by-nc-nd/3.0/es
info:eu-repo/semantics/openAccess
Reconeixement-NoComercial-SenseObraDerivada 3.0 Espanya (CC BY-NC-ND 3.0 ES)
oai:repositori.upf.edu:10230/463392021-02-05T02:30:58Zcom_10230_20650com_10230_16441col_10230_33097
2021-02-04T14:07:34Z
urn:hdl:10230/46339
Neural Tree-to-Tree Transduction
Morales, Cristian
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Leo Wanner, Bernd Bohnet
Graph-structured data is ubiquitous in the field of Natural Language Processing.
For instance, directed acyclic graphs are used in semantic and syntactic dependency
representations. Thus, several applications in NLP use graph-structured data, such
as sequence labeling, neural machine translation and relation extraction. Most approaches
first linearize graphs and then apply off-the-shelf algorithms, which leave
out important information of node connectivity. It is clear that the state-of-the
art falls short in offering graph-to-graph transduction models. The motivation of
this thesis is to expand the limited literature in tree-to-tree learning and provide
an instrument capable of treebank transformations, that could enrich the corpora
available in NLP. The starting point is previous work on Gated Graph Neural Networks
[1], which we modified to output sequences per node as opposed to sequences
per graph. We also modified the general architecture by using two GGNNs, one
responsible for predicting heads and the other one for predicting edge types. For
testing, we used the Stanford dependency treebank and the Matsumoto dependency
treebank. These treebanks are substantially different especially in the granularity of
their dependency tagsets. The proposed model achieved over 95% Labeled Attachment
Score (LAS) when converting from one treebank to the other. As compared
to the baseline, which ignores graph data, it achieved an average improvement of
16.42% in LAS, which highlights the value of incorporating graph-structured data.
We also showed that feeding the network with each node’s position within the sentence
yielded a 2.32% LAS improvement. Thus, including sequential data proved
to be beneficial. We concluded that GGNNs are capable of tree-to-tree transduction
and that this research is a step forward in bringing attention to graph-to-graph
transduction in NLP.
2021-02-04T14:07:34Z
2021-02-04T14:07:34Z
2020-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/46339
eng
https://creativecommons.org/licenses/by-nc-nd/3.0/es
info:eu-repo/semantics/openAccess
Reconeixement-NoComercial-SenseObraDerivada 3.0 Espanya (CC BY-NC-ND 3.0 ES)
oai:repositori.upf.edu:10230/463592021-02-06T02:30:43Zcom_10230_20650com_10230_16441col_10230_33097
2021-02-05T07:23:22Z
urn:hdl:10230/46359
Criteria for algorithmic fairness metric selection under different supervised classification scenarios
Breger, Clothilde
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Carlos Castillo
The research community, (supra-)national institutions, and regular users have noticed
that Artificial Intelligence and Machine Learning algorithms can amplify existing
inequity between groups. One way to limit this is to use group fairness metrics
to measure inequity, optimise and select models. However, there are many different
group fairness metrics. Here I combined a clustering of metrics (as done by Friedler
et al. in their 2019 paper "A comparative study of fairness-enhancing interventions
in machine learning" and by Miron et al. in their 2020 paper "Addressing multiple
metrics of group fairness in data-driven decision making") and expert-driven recommendations
(from a case study by Rodolfa et al., published in 2020: "Case study:
Predictive Fairness to Reduce Misdemeanor Recidivism Through Social Service Interventions")
to select fairness metrics. Although this clustering was not consistent,
it enabled fairness metric selection and fostered general recommendations on the
matter: an algorithm designer should extensively study their algorithm’s application
context and explicitly justify their choices relative to fairness. So long as there
is no absolute guide to metric selection, this should help nourish an ongoing and
context-specific discussion on algorithmic fairness, within and outside of the research
community.
2021-02-05T07:23:22Z
2021-02-05T07:23:22Z
2020-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/46359
eng
https://creativecommons.org/licenses/by-nc-nd/3.0/es
info:eu-repo/semantics/openAccess
Reconeixement-NoComercial-SenseObraDerivada 3.0 Espanya (CC BY-NC-ND 3.0 ES)
oai:repositori.upf.edu:10230/463602021-02-06T02:31:16Zcom_10230_20650com_10230_16441col_10230_33097
2021-02-05T07:31:14Z
urn:hdl:10230/46360
Generation of Deepfakes using normalizing flows
Valenzuela Ramírez, Andrea
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Vicenç Gómez, Ferran Diego Andilla, Carlos Segura Perales
Deepfakes are flooding the internet network with fake synthetic media that cannot
be easily distinguished by the human eye. This content can be classified into
face synthesis, face swap and facial attribute manipulation. Encoder-decoder and
generative networks are the leading architectures in deepfake creation due to the
fact that they produce very realistic and high quality content. Flow-based models
have been emerging in these recent years due to the several properties that make
them attractive for the scientific community. Nevertheless, only the techniques of
face synthesis and attribute manipulation have been explored with this architecture.
Here we present a first approach for face swap with flow-based models under the
assumption of transferring facial expressions between identities with vector arithmetic.
Different arithmetic expressions have been used to generate deepfake content
that has been later evaluated in terms of the likeliness to the target identity, the
quality of the expression transfer and the probability of the image of being fake.
The arithmetic expression that produces better deepfake content according to our
metric is the linear combination of the expression vector, obtained as the mid point
of the difference between the original expression of the source (with the expression
that wants to be transferred) and its mean face regulated by a factor α, and the
mean face of the target identity. The results show that there are no clear guidelines
on the best α value outputting a better expression transfer since this value
depends not only on the expression that wants to be transferred, but also on the
target identity. Regarding the quality of the synthesized expression, the obtained
mean error in terms of the intensity of the Action Units characterizing the selected
expressions is of 0:3 in the 0-5 intensity scale. This work supposes a first insight into
face swapping via expression transferring in flow-based models providing an initial
pipeline for both generation and evaluation of such type of deepfake content.
2021-02-05T07:31:14Z
2021-02-05T07:31:14Z
2020-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/46360
eng
https://creativecommons.org/licenses/by-nc-nd/3.0/es
info:eu-repo/semantics/openAccess
Reconeixement-NoComercial-SenseObraDerivada 3.0 Espanya (CC BY-NC-ND 3.0 ES)
oai:repositori.upf.edu:10230/463632021-02-06T02:30:54Zcom_10230_20650com_10230_16441col_10230_33097
2021-02-05T09:30:53Z
urn:hdl:10230/46363
Predicting multi-resistance of bacteria in an Intensive Care Unit
Hernàndez Carnerero, Àlvar
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Miquel Sànchez i Marrè, Vicenç Gómez
This study considers the prediction of “multi-drug” resistance (MDR) of Pseudomonas
aeruginosa bacterium caused by nosocomial infections in the Intensive
Care Unit (ICU). An ensemble of binary classifiers implemented with different Machine
Learning (ML) methods is applied for prediction using as training data health
records and past sensitivity tests (antibiogram) results. This work proposes to generate
two new types of features to improve predictor’s performance. The first one
is based on using information of previous antibiograms of a particular patient to
predict their future resistance to antibiotics. The second kind of features employs
bacterial information from the rest of the patients in the ICU to predict the antimicrobial
resistance for a certain patient. In addition, in the study it is suggested
to use a training window with incremental size so that training set is always temporarily
as near as possible to the test instances to be predicted. Some techniques
such as feature selection and oversampling are also used to further improve efficiency
and accuracy. Results show that using an incremental window for training improves
success rates in the domain of this problem, and expose that knowing the outcomes
of past antibiograms, substantially improves prediction. It is also observed that considering
resistant bacteria present in the ICU is useful to anticipate antimicrobial
resistance. From these results it is further inferred that resistant bacteria may be
spreading among patients in the ICU within populations that rapidly mutate, which
can induce non-stationary in the data distribution. It is concluded that using these
contributions, experiments show promising results in MDR prediction even using
simple features and limited training data.
2021-02-05T09:30:53Z
2021-02-05T09:30:53Z
2020-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/46363
eng
info:eu-repo/grantAgreement/ES/2PE/TIN2017-90567-REDT
https://creativecommons.org/licenses/by-nc-nd/4.0/
info:eu-repo/semantics/openAccess
Reconeixement-NoComercial-SenseObraDerivada 4.0 Internacional (CC BY-NC-ND 4.0)
oai:repositori.upf.edu:10230/465132021-02-19T02:31:00Zcom_10230_20650com_10230_16441col_10230_33097
2021-02-18T08:47:56Z
urn:hdl:10230/46513
Correlation of speech/non-speech events with photo-plethysmographic (PPG) signal
Cámbara Ruiz, Guillermo
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Jordi Luque, Mireia Farrús
The use of photoplethysmogram signal (PPG) for heart monitoring is commonly
found nowadays in smartphones and wrist wearables. Besides heart rate or sleep
monitoring common usage, it has been proved that information from PPG can be extracted
for other uses, like person verification, for example. In this work, we evaluate
whether if speech/non-speech events can be inferred from fluctuations they might
cause in the pulse signal. In order to do so, an exploration on end-to-end convolutional
neural network architectures is done for performing both feature extraction
and classification of the mentioned events. The results are motivating, detecting
speech in PPG signal with a 68.2% AUC using the best performing architecture.
On the other hand, a first experiment on speaker’s voice pitch detection is done, in
order to check if a prosody marker such as pitch variation could be present in PPGs,
but such clue is not clearly found in the results obtained. Nevertheless, the correlation
between speech and PPG signal is proven and the way is paved for further
experiments on this topic.
2021-02-18T08:47:56Z
2021-02-18T08:47:56Z
2019-06
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/46513
eng
https://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Reconocimiento-NoComercial-SinObraDerivada 3.0 España (CC BY-NC-ND 3.0 ES)
oai:repositori.upf.edu:10230/489492021-11-12T02:31:48Zcom_10230_20650com_10230_16441col_10230_33097
2021-11-11T07:41:39Z
urn:hdl:10230/48949
The bias effect of news media sources on social media users
Kavas, Hamit
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Leo Wanner
Bias in media news is such an interesting topic that it increases its popularity day by day.
Classification of news sources according to their political tendency is a very well studied
subject. However, we have not been able to find any research based on how these polit-
ical orientations are transferred to the users on social media. Thus, in this research we
aim to detect changes in the opinion of social media users over time in comparison to the
news articles. In light of this, we have created an hybrid model, by combining Convolu-
tional neural networks (CNN) with Long short-term memory (LSTM), which is initiated
by state-of-the-art BERT embeddings. Our choice of BERT embeddings was based on
a long trial and error process with other word embeddings such as GloVe, word2vec
and fastText, which led us to the conclusion that the transformer and attention mecha-
nism properties of BERT embeddings make it superior to the others. Domain adaptation,
which a transfer learning method is employed as the supplementary method in this pa-
per, to overcome the language usage differences between news articles and user tweets.
Thanks to transfer learning, we have observed a significant improvement in the models
performance. The created model has been trained on 600.000 labeled news articles and
84.000 politically leaned tweets during the transfer learning. As a result of our testing on
recently published and labeled news articles, as expected our model has been successful
in proving that politically biased news articles provoke Twitter users to make more biased
comments, whereas objective news articles do not lead people to express more political
bias. This thesis also includes an Apache NiFi implementation of the idea of monitoring
bias dynamically.
2021-11-11T07:41:39Z
2021-11-11T07:41:39Z
2020-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/48949
eng
http://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Atribución-NoComercial-SinDerivadas 3.0 España
oai:repositori.upf.edu:10230/492182021-12-16T02:31:58Zcom_10230_20650com_10230_16441col_10230_33097
2021-12-15T11:50:55Z
urn:hdl:10230/49218
Modelling the use of car parks in the province of Barcelona
Ferrer Sánchez, Josep
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Vicenç Gómez
Designing and generating wiser policies for urban systems and infrastructures is a
challenge of paramount importance. Today, the cities that present the most successful
transport strategies are prioritising the movement of people, giving residents
and visitors a wider variety of attractive transport options while creating effective
ways to switch from private to public transport means. Understanding the use and
the impact of public infrastructures that facilitate mobility is crucial.
We consider a dataset of one year of activity in the form of car park occupancy in
the province of Barcelona. The data comprises ten different park station facilities
located close to train stations. We propose and analyze different models based on
statistical and mathematical techniques.
First, we analyze the occupancy recordings in different park station locations and
show that the activity is strongly coupled with the circadian rhythm, following a
24-hours cyclic pattern. Second, we perform a statistical characterization based
on different activity prototypes for days associated with similar activity. We show
that such a simple statistical model is enough to characterize accurately the car
park occupancy. Third, we propose a mathematical model that explicitly captures
the rise and decay of occupancy using a mixture of two different Gamma kernels.
We optimize the parameters for each park station independently and show that the
resulting model globally captures better the activity than the previous statistical
characterization, in terms of accuracy, simplicity (has less parameters) and interpretability.
Our results show that, despite the apparent complexity associated to public mobility
and use of car parks, very simple models motivated in intuitive principles are
sufficient to understand this dynamics to a large extent. Overall, our results can
facilitate the design of public policies to facilitate the mobility within Barcelona and
its surroundings, by providing a better understanding of how the citizens switch
between private cars and public trains.
2021-12-15T11:50:55Z
2021-12-15T11:50:55Z
2021-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/49218
eng
info:eu-repo/semantics/openAccess
© Tots els drets reservats
oai:repositori.upf.edu:10230/492192021-12-16T02:31:59Zcom_10230_20650com_10230_16441col_10230_33097
2021-12-15T12:01:47Z
urn:hdl:10230/49219
Predicting the use of car parks in the province of Barcelona
Moreno Esteban, David
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Vicenç Gómez
Designing and generating wiser policies for urban systems and infrastructures is a
challenge of paramount importance. Today, the cities that present the most successful
transport strategies are prioritising the movement of people, giving residents
and visitors a wider variety of attractive transport options while creating effective
ways to switch from private to public transport means. Understanding the use and
the impact of public infrastructures that facilitate mobility is crucial.
We consider a dataset of one year of activity in the form of car park occupancy in
the province of Barcelona. The data comprises ten different parking facilities located
close to train stations. We propose and analyze different and intuitive prediction
models based on statistical and mathematical approximations.
First, we analyze the occupancy recordings in different parking locations and show
that the activity is strongly coupled with the circadian rhythm, following a 24-hours
cyclic pattern. Second, we implement a predictive model to provide the occupancy of
a particular parking for an entire future day. We show that for both, statistical and
mathematical approximations it performs quite accurately. Third, we implement a
predictive model to guess the occupancy of the remaining hours of the day given the
occupancy of the previous hours. Finally, a qualitative and quantitative analysis of
the parking occupancy during the Covid-19 pandemic has been performed in order
to understand how the global situation has influenced the parking usage.
Our results show that, despite the apparent complexity associated to public mobility
and use of car parks, very simple models motivated in intuitive principles are sufficient
to understand and predict this dynamics. Overall, our results can facilitate
the design of public policies to facilitate the mobility within Barcelona and its surroundings,
by providing a better understanding of how the citizens switch between
private cars and public trains.
2021-12-15T12:01:47Z
2021-12-15T12:01:47Z
2021-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/49219
eng
info:eu-repo/semantics/openAccess
© Tots els drets reservats
oai:repositori.upf.edu:10230/492212021-12-20T11:50:03Zcom_10230_20650com_10230_16441col_10230_33097
2021-12-15T12:11:10Z
urn:hdl:10230/49221
Selecting features of breathing & stability data to predict stress in students
Fischer, Johannes Simon
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Simone Tassani
Repeated or long-time exposure to psychological stress is a risk factor for mental and physical health. Especially students are exposed to stress and lacking stress management skills make them particularly vulnerable to stress-related illnesses. The Breathing Dynamic Modelling for Body Mind Interaction in Students (BYMBOS) project tries to tackle this problem by investigating the triangular relationship between breathing, posture, and psychological stress. This exploratory study contributes to this project by selecting discriminative physiological breathing and stability features that can detect stress. For this purpose, four feature selection methods were applied, namely a Genetic Algorithm, a Decision Tree, a Correlation-based, and a Relief-based feature selection approach. For the classification a Support Vector Machine (SVM), as well as the k-nearest Neighbor (KNN) classifier were used. Stability and respiratory data were recorded before and after relaxation was induced through deep breathing and muscle relaxation exercises. The application of the feature selection methods on the physiological data confirmed that including feature selection as a preprocessing step can not only reduce the number of variables significantly but can also increase accuracy, due to the elimination of noise. Predicting stress based on breathing and stability data achieved a leave-one-subject-out (LOSO) mean accuracy of up to 85.42%. Stability measures were found to boost the predictive power but were not sufficient for an effective prediction. However, these results may be biased as the relaxation inducing intervention included deep breathing exercises and several subjects were not able to perform deep breathing correctly. Hence, a different labelling procedure based on whether a person was able to perform deep breathing or not, was proposed. The prediction with this labelling solely based on stability measures achieved a LOSO mean accuracy of up to 87.76%. This suggests a relationship between stability, breathing, and psychological stress. Nonetheless, further research with more samples, and procedures to validate the perceived stress in subjects is required to support this hypothesis.
2021-12-15T12:11:10Z
2021-12-15T12:11:10Z
2021-07
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/49221
eng
https://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Attribution-NonCommercial- NoDerivs 3.0 Spain
oai:repositori.upf.edu:10230/492222021-12-16T02:32:01Zcom_10230_20650com_10230_16441col_10230_33097
2021-12-15T12:17:43Z
urn:hdl:10230/49222
Learning non-linear payoff transformations in multi-agent systems
Fraxanet, Emma
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Vicenç Gómez
The use of Deep Reinforcement Learning methodologies has been successful in recent
years in cooperative multi-agent systems. However, this success has been mostly empirical and there is a lack of theoretical understanding and solid description of the
learning process of those algorithms. The discussion of whether the limitations of
these algorithms can be tackled with tuning and optimization or, contrarily, are constrained
by their own definition in these models can also easily be put forward. In
this work, we propose a theoretical formulation to reproduce one of the claimed limitations
of Value Decomposition Networks (VDN), when compared to its improved
related model QMIX, regarding their representational capacity. Both of these algorithms
follow the centralized-learning-decentralized-execution fashion. For this
purpose, we scale down the dimensions of the system to bypass the need for deep
learning structures and work with a toy model two-step game and a series of one-shot
games that are randomly generated to produce non-linear payoff growth. Despite
their simplicity, these settings capture multi-agent challenges such as the scalability
problem and the non-unique learning goals. Based on our analytical description, we
are also able to formulate a possible alternative solution to this limitation through
the use of simple non-linear transformations of the payoff, which sets a possible
direction of future work regarding larger scale systems.
2021-12-15T12:17:43Z
2021-12-15T12:17:43Z
2021-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/49222
eng
info:eu-repo/semantics/openAccess
© Tots els drets reservats
oai:repositori.upf.edu:10230/492232021-12-20T13:26:20Zcom_10230_20650com_10230_16441col_10230_33097
2021-12-15T12:25:34Z
urn:hdl:10230/49223
Thompson sampling for prediction with expert advice
Sánchez Pérez de Amézaga, Claudio
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Gergely Neu
We study Thompson Sampling for prediction with expert advice. With Follow
the leader, and Follow the perturbed leader strategies, we present relevant results in
order to proceed with Thompson Sampling Algorithm. Using a similar strategy used
for studying Follow the perturbed leader, we decompose the regret in three terms. We
study the expressions of choosing an expert from a set of experts. Here we show
some interesting equivalences between the probability at time t, and the probability
of a cheating forecaster which can see in the future t + 1. Finally, we present some
experimental cases xing a nal time T. We analyze how the model selects the expert
with the best performance. We obtain strong evidence to bound the diference between
the cheating forecaster and the true one, following an exponential growth similar to
T.
2021-12-15T12:25:34Z
2021-12-15T12:25:34Z
2021-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/49223
eng
https://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Attribution-NonCommercial- NoDerivs 3.0 Spain
oai:repositori.upf.edu:10230/492242021-12-16T02:32:03Zcom_10230_20650com_10230_16441col_10230_33097
2021-12-15T12:31:56Z
urn:hdl:10230/49224
Multilingual lexical simplification
Pimienta Castillo, Jorge S.
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutor: Horacio Saggion
This report describes, implement, and evaluate one strategy for text simplification,
namely, Lexical Simplification, that aims to reduce the complexity of some words
in a sentence. This process is done in two main steps, the first, is a module that
identifies the complex elements, and the second, is a module that replaces those
elements for simpler variants.
For the first module, the system will use three different datasets that include human
annotations in different languages: English, Spanish, and German, this will allow us
to train a classifier that detects complex words.
For the second module, a pre-trained model for word prediction (BERT) will be used
to generate the candidates, the candidates will be sorted based on Zipf’s frequency,
to later select the one with the highest value.
Finally, the complete system is evaluated using a test dataset, and a survey designed
to collect human annotations and perception of Fluency, Meaning and Simplicity.
2021-12-15T12:31:56Z
2021-12-15T12:31:56Z
2021-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/49224
eng
https://creativecommons.org/licenses/by-nc-nd/4.0/
info:eu-repo/semantics/openAccess
Attribution-NonCommercial- NoDerivs 4.0 International
oai:repositori.upf.edu:10230/554502023-01-28T02:31:56Zcom_10230_20650com_10230_16441col_10230_33097
2023-01-27T18:27:48Z
urn:hdl:10230/55450
Linguistic emergence in deep neural networks
Locatelli Metz, Davide
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Marco Baroni, Roberto Dessi
Recent work in Emergent Communication has shown that deep neural networks can develop successful
strategies of referential communication. Arguably, the extent to which they can do so
depends on the properties of the communication channel that is provided to them. Here, we test
this hypothesis by analyzing the protocols emerging from visual referential games with increasingly
complex communication channels. More specifically, we increase the complexity by varying the
vocabulary size and the length of the messages exchanged. We show that this directly influences
communication success and profoundly impacts the chosen referential strategy. Our results evidence
a tendency of deep nets to performbetter with greater vocabulary sizes. Moreover, we found
that deep nets prefer symbolic communication, i.e. when messages consist of single symbols, rather
than combinatorial languages, i.e. when messages consist of sequences of two or more symbols. Finally,
we showthat the most successful communication strategy is achieved when the nets converge
to a protocol that approximates a one-to-one mapping between messages and images
2023-01-27T18:27:48Z
2023-01-27T18:27:48Z
2022
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/55450
eng
http://creativecommons.org/licenses/by-nc-nd/4.0
info:eu-repo/semantics/openAccess
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License
oai:repositori.upf.edu:10230/554512023-01-28T02:31:57Zcom_10230_20650com_10230_16441col_10230_33097
2023-01-27T18:56:09Z
urn:hdl:10230/55451
A novel message passing approach to spatial air quality prediction in urban areas
Calo Oliveira, Sergio
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Filippo Bistaffa, Vicenç Gómez, Anders Jonsson
Air pollution in our cities is a very significant cause of death and worsening of
the quality of life in current times. Knowing in depth this phenomenon and its
details and building tools that help us to mitigate its effects can be key to a better
habitability of present and future urban environments. This is precisely the purpose
of this work. Using specific data from the city of Barcelona, the problem of spatial
prediction of air quality at the urban level has been addressed. The objective is
to obtain a resolution at street level starting from the known values of a series of
detectors, dispersed throughout the city. Certain particularities of this case study
have forced us to carry out novel research in this aspect. The contributions of
this work are the following. Firstly, a formalization of the problem based on Markov
Random Fields has been carried out, following the recommendations of recent works
in this sense, in order to favor the unification of the theoretical frameworks of the
Graph Signal Processing field. After this, a novel algorithm has been developed for
the resolution of the problem. This algorithm is based on a message passing scheme
between nodes. In the proposed method, this algorithm is combined with a Graph
Neural Network that refines the obtained result for a better approximation.
2023-01-27T18:56:09Z
2023-01-27T18:56:09Z
2022
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/55451
eng
https://creativecommons.org/licenses/by-nc/3.0/es/
info:eu-repo/semantics/openAccess
Reconocimiento-NoComercial 3.0 España (CC BY-NC 3.0 ES)
oai:repositori.upf.edu:10230/554522023-01-28T02:31:58Zcom_10230_20650com_10230_16441col_10230_33097
2023-01-27T19:05:53Z
urn:hdl:10230/55452
Task and motion planning: scaling up
Dalmau Moreno, Magí
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Héctor Geffner, Néstor García
This master thesis contributes to provide planning capabilities to embodied agents so
that they are able to decide what sequence of actions should be taken to solve a given
complex problem, involving coupled reasoning both at symbolic and geometric level
(i.e. combined task and motion planning). This work has investigated width-based
search algorithms, and a new algorithm has been proposed which implements a novel
lazy-interleaved approach to reduce the computational cost of the planning, without
making any assumptions on the robot used or the environment objects. Note,
then, that the proposed approach works completely online without needing any precomputation
at all. This work effectively exploits domain knowledge provided via
sketches to guide the search in problems with huge-combinatorial state-spaces. The
proposed approach is able to work transparently with robotic continuous search state
spaces with flexibility, using an adaptive sampling of the world to build the search
space of the problem. Moreover, the problem dynamics are retrieved as a black
box, so the developed planner is able to work directly with a simulator, and it does
not need an explicit declaration of the action structure. The proposed approach
has been validated in two problem families that illustrate the current challenges
in combined task and motion planning (video demos showing plans calculated using
the developed framework can be found in https://drive.google.com/drive/
folders/10goVJ8A86RGIGsbthdmak8L_mTdT6hby?usp=sharing). Furthermore, the
proposed approach has been compared with state-of-the-art approaches, obtaining
significantly better results. Besides, the proposed approach has been combined and
integrated within the ROS environment in order to provide the highest level of
standardization and compatibility with the maximum number of robotic systems
currently available.
2023-01-27T19:05:53Z
2023-01-27T19:05:53Z
2022
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/55452
eng
info:eu-repo/semantics/openAccess
©Tots els drets reservats
oai:repositori.upf.edu:10230/554532023-01-28T02:32:00Zcom_10230_20650com_10230_16441col_10230_33097
2023-01-27T19:17:08Z
urn:hdl:10230/55453
Taxonomic classification of metagenomic reads using machine learning models
Alcobé Garcia, Eduard
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Mario Ceresa, Antonio Puertas, Vicenç Gómez
Microorganisms such as bacteria can be hard to identify correctly. Most current
classification techniques are based on well conserved genes, for instance the 16S ribosomal
RNA (16S rRNA). Nevertheless, achieving a classifier with high accuracy
in classifying bacteria through 16S rRNA data is still a challenge. For this reason,
different machine learning approaches exploring a k-mer representation technique
can still contribute to solve this problem. Mapping the DNA sequences as vectors
in a numerical space, by counting the frequency of each k-mer in a given sequence,
is essential to be able to train the machine learning algorithms. Two deep learning
models, Convolutional Neural Networks and Deep Belief Networks, as well as
a tree-based model, XGBoost, are trained with synthetic datasets with 16S rRNA
sequences. These synthetic datasets are the 16S rRNA shotgun (SG), and the amplicon
(AMP) which considers only specific 16S hypervariable regions. Comparing
the performance of these models with the synthetic datasets provides useful information.
Moreover, it is also relevant to explore how these models work with real data
available in public genomic databases (NCBI, SILVA and FDA-ARGOS). Analysing
the classifiers’ performance with real data contributes to give an estimation of the
reliability of both, classifiers and public genomic databases.
2023-01-27T19:17:08Z
2023-01-27T19:17:08Z
2022
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/55453
eng
info:eu-repo/semantics/openAccess
©Tots els drets reservats
oai:repositori.upf.edu:10230/555042023-02-01T02:32:39Zcom_10230_20650com_10230_16441col_10230_33097
2023-01-31T18:45:59Z
urn:hdl:10230/55504
Extraction and categorization of Japanese lexical collocations with graph-aware transformers
Nisho, Kosuke James
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Leo Wanner, Alexander Shvets
Lexical collocations may be identified and categorized in context, which is helpful
for language acquisition, dictionary creation, and many other downstream NLP
tasks. However, the automatic collocation extraction and categorization using modern
machine learning techniques has not been tried in Japanese. In this paper, a
previous work in context-sensitive collocation identification using a sequence tagging
BERT-based model improved with a graph-aware transformer architecture is used
to investigate its feasibility to Japanese Language. The findings provide the initial
insights into the automatic collocation typification in a non Indo-European language
using deep learning models, and suggests that low resource languages can benefit
from this approach.
2023-01-31T18:45:59Z
2023-01-31T18:45:59Z
2022
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/55504
eng
http://creativecommons.org/licenses/by-nc-nd/4.0
info:eu-repo/semantics/openAccess
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License
oai:repositori.upf.edu:10230/555052023-02-01T02:32:41Zcom_10230_20650com_10230_16441col_10230_33097
2023-01-31T19:06:46Z
urn:hdl:10230/55505
Inverse optimal control for modeling virus mutations in SARS-CoV-2
Meijer, Ilse
Treball fi de màster de: Master in Intelligent Interactive Systems
Tutors: Vicenç Gómez, Mario Ceresa, Antonio Puertas Gallardo
Inverse Optimal Control (IOC) deals with the problem of recovering an unknown
cost function in a Markov decision process from expert demonstrations acting optimally.
In this thesis we apply IOC to SARS-CoV-2 data. For our application we use
the (mutated) sequences found in SARS-CoV-2 data as the expert demonstrations.
We present a way to learn useful state representations for this data, and successfully
apply IOC on a special class of Markov decision processes which allow for an
efficient computation of the value and cost functions of the states, and informative
2D representations of the state.
2023-01-31T19:06:46Z
2023-01-31T19:06:46Z
2022
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/55505
eng
http://creativecommons.org/licenses/by-nc-nd/4.0
info:eu-repo/semantics/openAccess
ement-NoComercial-SenseObraDerivada 4.0 Internacional →
dc.rights This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License
oai:repositori.upf.edu:10230/580412023-10-09T13:36:42Zcom_10230_20650com_10230_16441col_10230_33097
2023-10-04T17:34:01Z
urn:hdl:10230/58041
Predictive gut microbiome analysis for health assessment
Gil Sorribes, Manel
Treball fi de màster de: Master in Intelligent Interactive Systems.
Tutors: Mario Ceresa, Gabriele Leoni, Antonio Puertas, Mauro Petrillo
Personalized medicine is a rapidly evolving field to which many resources have been
devoted recently. It represents a paradigm shift from a one-size-fits-all approach
to healthcare, focusing instead on tailoring treatments and diagnoses to individual patients. This study aims to contribute to this transition by leveraging recent
advancements in microbiome research.
An earlier study that computed the Gut Microbiome Health Index (GMHI), a potent
indicator capable of predicting disease presence with approximately 73% of accuracy,
serves as the foundation for this research. The GMHI is a numerical value that
classifies a person as healthy if the index is greater than zero, non-healthy if less
than zero, and undetermined if the GMHI is equal to zero. The intent of this study
is to further utilize and develop this index through two primary objectives.
The first objective is to delve deeper into the dataset provided by the original GMHI
study authors, using statistical and Machine Learning (ML) techniques. This process involves a variable analysis, that are species, and the application of classifier
algorithms such as the Support Vector Machine. Upon determining the most informative variables, these will be used as input to a Neural Network (NN) in an
attempt to surpass the existing GMHI prediction accuracy. Through leveraging the
sophisticated pattern recognition capabilities of a NN, the aim is to refine the classification accuracy of health statuses, thereby providing more personalized diagnostic
insights. Two different architectures were used: a Fully Connected NN and an Autoencoder NN. The results obtained from this first objective, though not exceptional,
are promising. The accuracy slightly improved by approximately 3 percent when
using both approaches, indicating better prediction of health conditions.
The second objective is to apply the GMHI to a distinct dataset, which comprises
information on individuals both affected and unaffected by COVID-19. This novel
application of the GMHI represents an innovative investigation into the potential
impact of a viral disease like COVID-19 on an individual’s gut microbiome. Given
the worldwide repercussions of COVID-19, such exploration is crucial. As the potential implications of COVID-19 on the gut microbiome are largely unknown, this
research could provide key insights, influencing patient treatment strategies and
prognosis. Encouraging findings demonstrate that by rescaling and adapting the
GMHI to this specific dataset, the accuracy in detecting a COVID-19 patient significantly improves by 17 percent, thereby achieving an accuracy of 75%, compared
to the accuracy attained when directly using the authors’ formula.
In summary, this study aims to integrate advanced ML techniques to enhance the
GMHI and apply it to a novel dataset. This could significantly contribute to the burgeoning field of personalized medicine, where unique microbiome profiles may inform
diagnostic and therapeutic strategies. The integration of computational techniques
like ML could revolutionize the understanding of disease pathogenesis and treatment
approaches, propelling healthcare into a new era. This project findings may aid in
this direction and could serve as a baseline for further research, as outlined in the
further steps section.
2023-10-04T17:34:01Z
2023-10-04T17:34:01Z
2023-10-04
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/58041
eng
info:eu-repo/semantics/openAccess
Tots els drets reservats
oai:repositori.upf.edu:10230/580632023-10-10T01:30:24Zcom_10230_20650com_10230_16441col_10230_33097
2023-10-09T13:40:26Z
urn:hdl:10230/58063
Inverse reinforcement learning with linearly-solvable MDPs for multiple
reward functions
Deb, Ahana
Treball fi de màster de: Master in Intelligent Interactive Systems. Tutors: Anders Jonsson, Vicenç Gómez, Mario Ceresa
A subclass of Markov Decision Processes (MDPs), the Linearly solvable Markov
Decision Processes (LMDPs), which have discrete state space and continuous control
space, allow for a significant simplification of the inverse reinforcement learning
problem by eliminating the need to solve the forward problem, and requiring only the
unconstrained optimization of a convex and easily computable log-likelihood. This
however, has only been explored for the single-reward single-agent scenario, where a
single agent is assumed to be imposing optimal control under the influence of a single
fixed reward function. In this work, we aim to utilise the advantages in problem
formulation and ease of computation for LMDPs, for a multiple-agent, multiple-
reward scenario, using non-parametric Bayesian inverse reinforcement learning.
2023-10-09T13:40:26Z
2023-10-09T13:40:26Z
2023-10-09
info:eu-repo/semantics/masterThesis
http://hdl.handle.net/10230/58063
eng
https://creativecommons.org/licenses/by-nc-nd/3.0/es/
info:eu-repo/semantics/openAccess
Attribution-NonCommercial- NoDerivs 3.0 Spain