Doctoral School

Reading date: 13/02/2026

• FERRANDO MONSONÍS, JAVIER: Interpretability in Natural Language Processing and Machine Translation
  
  Author: FERRANDO MONSONÍS, JAVIER
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN ARTIFICIAL INTELLIGENCE
  Department: Department of Computer Science (CS)
  Mode: Normal
  Deposit date: 13/11/2025
  Reading date: 13/02/2026
  Reading time: 12:00
  Reading place: Sala MERIT D5010, Edifici D5, Campus Nord UPC, Barcelona
  Thesis director: RUIZ COSTA-JUSSA, MARTA
  Thesis abstract: This thesis presents a set of methods and analyses designed to improve our understanding of the internal mechanisms of Transformer-based models in natural language processing and machine translation.This work first investigates the role of attention weights in encoder-decoder Transformers, showing that while they do not provide accurate word alignments, they nonetheless help explain model predictions and contribute to a deeper understanding of translation quality.A central contribution of the dissertation is the development of ALTI and its extensions, which offer a new approach to input attribution. These methods challenge prior assumptions about the explanatory power of attention mechanisms and reveal how information propagates between encoder and decoder components. In doing so, they also shed light on sources of hallucinations in translation systems.Further, the thesis introduces techniques to attribute predictions to individual components and positions, enabling contrastive explanations of linguistic behavior. These explanations clarify how language models represent and solve different linguistic phenomena.The dissertation also proposes a methodology for tracking information flow during inference, offering insight into how various components contribute to model predictions. This allows for the identification of domain-specialized components and a better understanding of how representations are transformed across layers.Lastly, the analysis of cross-lingual circuit similarities reveals shared structural patterns in how models handle different languages. These findings point to potential universal mechanisms in language models.Collectively, this thesis advances the interpretability of Transformer models by providing tools and frameworks for probing, attributing, and understanding the behavior of complex NLP systems.
  

• PÉREZ GUIJARRO, JORDI: On Quantum Supervised Learning and Learning Techniques for Quantum Error Mitigation
  
  Author: PÉREZ GUIJARRO, JORDI
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN SIGNAL THEORY AND COMMUNICATIONS
  Department: Department of Signal Theory and Communications (TSC)
  Mode: Normal
  Deposit date: 22/12/2025
  Reading date: 13/02/2026
  Reading time: 11:00
  Reading place: Aula D5-007, Edifici D5, Campus Nord UPC, Barcelona
  Thesis director: RODRIGUEZ FONOLLOSA, JAVIER | PAGES ZAMORA, ALBA MARIA
  Thesis abstract: The development of quantum computers promises to drastically reduce the time required to solve certain computational problems. Among their most promising applications is the field of machine learning. However, significant uncertainty remains in this area. In particular, it is still unclear under which learning scenarios quantum algorithms will outperform their classical counterparts. This thesis aims to deepen our understanding of when quantum speed-ups can be expected in machine learning tasks. Specifically, we examine the connection between learning speed-ups and the more extensively studied phenomenon of quantum computational speed-up. We conclude that, in cases where the training set can be classically generated, the two are equivalent concepts, and we provide examples of such functions based on the prime factorization problem.Importantly, quantum machine learning is not only concerned with improving classical learning algorithms using quantum computation but also with learning from quantum data. In this context, we investigate a learning scenario in which the inputs to the target functions are quantum states, thereby generalizing the classical supervised learning framework. To this end, we first focus on the problem of quantum hypothesis testing, which can serve as a subroutine for both the problems of evaluating a function and learning a function. Specifically, we derive several sequential methods for solving the problem of quantum hypothesis testing, along with a lower bound on the resources required. This lower bound immediately implies corresponding lower bounds for the problems of learning and evaluating functions. Additionally, we develop a learning method based on the classical shadows technique.Finally, after exploring how quantum processes can aid learning tasks, we examine how classical learning techniques can, in turn, enhance quantum computing. In particular, we study how classical machine learning methods can be used to mitigate the effects of noise in quantum devices, with a focus on quantum error mitigation. Specifically, novel feature maps are proposed for the technique known as Clifford data regression. First, a theoretical justification for these feature maps is provided, followed by an analysis and a subsequent evaluation of their performance through numerical experiments. It is concluded that, for some of the proposed feature maps, a performance improvement is indeed achieved.
  

Reading date: 18/02/2026

• BORJA ROBALINO, RICARDO STALIN: Optimización bayesiana en técnicas machine Learning clásicas: redes neuronales y XGBoost y su aplicación como modelos predictores de diabetes en pacientes ecuatorianos
  
  Author: BORJA ROBALINO, RICARDO STALIN
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN STATISTICS AND OPERATIONS RESEARCH
  Department: Department of Statistics and Operations Research (EIO)
  Mode: Normal
  Deposit date: 22/01/2026
  Reading date: pending
  Reading time: pending
  Reading place: pending
  Thesis director: MONLEON GETINO, ANTONIO | GIBERT OLIVERAS, CARINA
  Thesis abstract: Machine learning (ML) is a branch of artificial intelligence that allows human capabilities to be imitated through various algorithms and techniques that learn from data using learning processes (supervised, unsupervised, or reinforcement) for decision-making with minimal human intervention. Classic ML models have generated great results in the automation of classification and regression processes in various areas. Within classification, artificial neural networks (ANN) have gained relevance due to their ability to learn and model complex nonlinear relationships. Similarly, the XGBoost model based on decision trees has demonstrated great efficiency, speed, scalability, and performance, winning several competitions. On the other hand, Bayesian inference has provided a probabilistic and revolutionary framework for optimizing machine learning models, with the implementation of uncertainty in the estimation process, combining evidence with prior beliefs, in order to reduce overfitting and improve predictions by adjusting parameters and hyperparameters.This research aims to optimize two classic machine learning techniques (artificial neural networks and XGBoost) for classification using Bayesian inference and to build a diabetes prevention model for the Ecuadorian population. The study begins with a theoretical and mathematical conceptualization of each algorithm, followed by an analysis of the various points of intervention, programming, and implementation of Bayesian models using Markov chain Monte Carlo (MCMC) estimation techniques and variational inference (VI), validation using public databases, implementation of a client-server system with multiple specialized backends, and, finally, the development of a real application as predictors of type 1, type 2, and gestational diabetes.As a result, a Bayesian model was implemented in artificial neural networks (ANN) at two inference points. The first adjusted the parameters at each backpropagation step; however, it presented itself as an option with a prohibitive computational overhead. As a second intervention, an adjustment was made to the activation function in the final layer, obtaining positive and computationally viable results. In the case of XGBoost, the predictions were adjusted at each boosting step before vectorization, demonstrating high predictive power in both the use of the MCMC technique and IV. Validation with the Pima Indians Diabetes database and the use of various distribution functions demonstrated the robustness and sensitivity of the implemented models, while generalization and consistency were verified through application to various databases. In all cases, results superior to or equal to those obtained using the traditional model were obtained, depending on the characteristics of the data.In addition, a web application (client-server) was implemented with Bayesian proposals, allowing users to interact with the models in an easy and intuitive way, with options for data loading, parameter configuration and probability distributions, estimation techniques (MCMC or IV), training-validation process or use of cross-validation, real-time results, and model download options. The application of the Bayesian proposal to a real case, such as the prediction of type 1, type 2, and gestational diabetes, with data from Ecuadorian patients, presented encouraging results (accuracy = 99.47%), becoming the first predictive model for the three types of diabetes at the regional and national level, confirming that the use of this approach is an excellent alternative for the optimization of machine learning models.
  

• CASTRO CARRASCO, REBECA IGNACIA: Characterization of sulfate-reducing biofilms using an amperometric printed H₂S sensor
  
  Author: CASTRO CARRASCO, REBECA IGNACIA
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN NATURAL RESOURCES AND THE ENVIRONMENT
  Department: Department of Mining, Industrial and ICT Engineering (EMIT)
  Mode: Normal
  Deposit date: 22/01/2026
  Reading date: pending
  Reading time: pending
  Reading place: pending
  Thesis director: GABRIEL BUGUÑA, GEMMA | GUIMERÀ VILLALBA, XAVIER
  Thesis abstract: A comprehensive understanding of sulfidogenic processes in bioreactors remains incomplete by the limited availability of tools suitable for the sulfate-reducing activity characterization of immobilized biomass. To address this limitation, the present work is based on developing suitable alternatives for sulfate reducing biomass characterization using electrochemical microsensors. In this sense, a flow-cell bioreactor was developed for real-time monitorization using artificially immobilized biomass to substitute the natural immobilization derived from extracellular polymeric compounds. Physical and functional evaluations enabled the identification of a polymer–biomass matrix capable of preserving sulfate-reducing performance while ensuring adequate microbial retention and structural integrity, as well as a range of operational conditions was assessed to generate detailed H₂S production profiles within the flow-cell bioreactor. In parallel, an inkjet-printed H₂S microsensor was fabricated on polyethylene terephthalate substrates using silver and gold inks and modified with Single-walled carbon nanotubes reinforced with Polyvinyl alcohol and Polydiallyldimethylammonium chloride, which improved ink dispersion, adhesion, and mechanical stability. The optimized formulation yielded long-term operational stability, linear responses across different media, and performance comparable to commercial microsensors despite an initial decrease in sensitivity. Furthermore, the study evaluated the operational behavior of artificial sulfate-reducing granules in column and continuous stirred tank reactors, demonstrating high sulfate removal efficiencies at moderate loading rates, superior stability in column configurations, accumulation of volatile fatty acids associated with incomplete glycerol oxidation, and the effectiveness of a bioaugmentation strategy based on acetate-oxidizing sulfate reducing bacteria immobilized in artificial granules. Lastly, the integrated platform was validated for the analysis of H₂S production in immobilized sulfate-reducing biofilms, combining the flow-cell bioreactor with direct ink writing printed microsensors for simultaneous, in situ monitoring of H₂S and pH. Three-dimensional mapping revealed pronounced H₂S gradients driven by mass-transfer limitations and hydrodynamic dispersion, while printed electrodes exhibited linear amperometric responses and stable performance over extended operation, thereby confirming the suitability of the proposed platform for high-resolution, real-time characterization of sulfidogenic biofilms and immobilized sulfate-reducing biomass.
  

• MUÑOZ GALAN, HELENA: Sensor design and development for autonomous devices for disease diagnosis and therapy
  
  Author: MUÑOZ GALAN, HELENA
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN POLYMERS AND BIOPOLYMERS
  Department: Department of Chemical Engineering (EQ)
  Mode: Normal
  Deposit date: 22/01/2026
  Reading date: pending
  Reading time: pending
  Reading place: pending
  Thesis director: ALEMAN LLANSO, CARLOS ENRIQUE | PÉREZ MADRIGAL, MARIA DEL MAR
  Thesis abstract: This doctoral research addresses key challenges in diabetes management by integrating sustainable materials, non-invasive sensing, and advanced insulin delivery technologies into a unified framework. Diabetes mellitus is a chronic metabolic disorder that requires continuous glucose monitoring and precise insulin administration to maintain glycemic control and reduce complications. Responding to these needs, the thesis is structured around three major contributions.First, the work advances the development of a previously created non-invasive glucose-monitoring device by incorporating recycled low-density polyethylene (LDPE) into the sensor design. The use of recycled LDPE improves sustainability, cost-efficiency, and environmental compatibility while maintaining reliable sensing performance. This demonstrates the feasibility of repurposing plastic waste in biomedical technologies without compromising functionality.Second, the research explores new strategies for controlled insulin delivery through stimuli-responsive hydrogels. Hydrogels based on poly(γ-glutamic acid) and multi-armed polyethylene glycol (PEG) were engineered to achieve sustained and tunable insulin release. The incorporation of poly(3,4-ethylenedioxythiophene) (PEDOT), a biocompatible conducting polymer, enables electrically triggered and on-demand insulin delivery. This approach offers a minimally invasive alternative to conventional injection-based therapies and highlights the potential of electro-responsive materials in smart drug-delivery systems.Finally, the thesis investigates the nanomechanical properties of multi-armed PEG hydrogels using a microcantilever-based optomechanical sensor. This analysis provides essential insights into the structural behavior, stability, and long-term performance of the hydrogels, contributing to the optimization of their mechanical robustness and responsiveness.Together, these contributions form a comprehensive strategy for improving diabetes care. By integrating non-invasive monitoring, environmentally conscious materials, and intelligent delivery platforms, the research promotes more sustainable biomedical solutions while advancing patient-centered therapeutic tools.