Development of neural networks by ordinal, explainable, and multi-task learning: applications in biomedicine and clean energies

• Autores: ANTONIO MANUEL GOMEZ ORELLANA
• Directores de la Tesis: Pedro Antonio Gutiérrez Peña (dir. tes.), David Guijo Rubio (dir. tes.)
• Lectura: En la Universidad de Córdoba (ESP) ( España ) en 2025
• Idioma: inglés
• Tribunal Calificador de la Tesis: Ezequiel López Rubio (presid.), Nicolás García-Pedrajas (secret.), Rafaela Benitez Rochel (voc.)
• Programa de doctorado: Programa de Doctorado en Computación Avanzada, Energía y Plasmas por la Universidad de Córdoba
• Enlaces
  • Tesis en acceso abierto en: Helvia
• Resumen

  • Artificial Intelligence (AI) research is dedicated to the study and development of systems that are capable of performing tasks autonomously. These AI-based systems were initially scarce; however, thanks to the continuous advances in programming languages and in computing power, as well as the availability of data from many sources, these systems have become part of everyday life. A particular branch of AI that has contributed greatly to this spread and success is Machine Learning (ML), which focuses on making computers learn from examples. Among ML techniques, Artificial Neural Networks (ANNs) have shown an outstanding performance. Specifically, ANNs are applied to solve a variety of real-world problems in diverse fields such as healthcare, renewable energy, and education, among others, thereby contributing to innovation in our daily lives. However, distinct challenges arise when approaching certain tasks using ANNs, some of which will be addressed in this Thesis.

    Multi-Task Learning (MTL) is a paradigm in which ANNs are an excellent approach thanks to their versatile architecture. This paradigm leverages information from related tasks, which are learned concurrently, with the aim of enhancing performance while reducing complexity. However, it is challenging to determine which ANN architecture would be more appropriate for a given MTL problem. Thus, this Thesis explores the learning of ANN architectures in MTL problems using evolutionary computation.

    ANNs are also a valuable technique in the field of Ordinal Classification (OC), a specific type of nominal classification in which the classes are ordered. Two common strategies for OC using ANNs are: 1) incorporating an ordinal output layer; and 2) using an ordinal loss function. Based on both strategies, this Thesis proposes two methods for OC with ANNs.

    While ANNs have shown remarkable performance, their non-linear structure often obscures the internal mechanisms by which they solve specific tasks. The lack of interpretability is a significant obstacle for ANNs in certain areas, limiting their adoption. Therefore, this Thesis considers an eXplainable Artificial Intelligence (XAI)-based approach to ANNs learning, with the aim of improving their explainability and interpretability.

    Survival analysis is a specific task that involves analysing the time until a certain event occurs, and it is typically approached using linear statistical methods. However, ANNs can model non-linearities, which are often found in survival analysis data, producing more accurate results than linear models. Therefore, this Thesis explores the use of ANNs in survival analysis from an XAI perspective.

    Finally, the contributions presented in this Thesis are validated in a wide range of real-world problems, including the following fields of application: clean energies, biomedicine, and meteorology.