Interests
Neural Networks and Learning Models
Can computers autonomously learn to solve problems based on examples? This has been one of the fundamental questions driving my research in the field of artificial neural networks. One of the most fascinating properties of these networks is their ability to generalize to new inputs, applying acquired knowledge to previously unseen situations. This phenomenon is crucial for developing computational models capable of efficiently automating problems that may initially seem highly complex.
In collaboration with my thesis advisor, Leonardo Franco, I have dedicated significant time to deeply analyze the computational capabilities of various neural network architectures. Our objective has been to develop innovative methods that not only enhance the generalization capability of these networks, but also provide insights into how this process operates in practice. One of the most significant challenges in using artificial neural networks for learning is the uncertainty surrounding the optimal network topology. To overcome this obstacle, regularization techniques and exhaustive search methods are typically employed to discover effective and validated architectures.
My most notable contribution has been in the development of constructive learning algorithms. These algorithms not only enable the automatic construction of nearly optimal neural architectures, but also facilitate knowledge extraction, offering valuable insights into how networks solve specific problems.
Constructive Neural Networks represent an innovative approach to addressing the challenge of selecting the most suitable neural architecture for a given problem. In this context, we have developed two key algorithms: DASG and C-Mantec. DASG is a decomposition algorithm that analyzes a Boolean function (or data) to decompose it into linearly separable functions based on the influence of its variables. On the other hand, C-Mantec is a highly powerful constructive neural network algorithm that utilizes competition between neurons to create compact architectures with excellent generalization capabilities.
In addition to applying these algorithms in various academic and research contexts, we have explored their practical application in Wireless Sensor Networks. In collaboration with Francisco Ortega, we have successfully implemented the C-Mantec algorithm on FPGA devices, enabling extremely fast and efficient execution.
Currently, the field of neural networks is being transformed by advances in deep learning algorithms. This approach, known as Deep Learning, represents an evolution from traditional back-propagation neural networks by incorporating architectures with multiple hidden layers. The results achieved with Deep Neural Networks to date have been impressive, surpassing human capabilities in solving complex problems in several cases.
My ongoing interest in this research area is focused on continuing to develop constructive neural network models, especially within the realm of deep learning, where the prospects for advancement and practical application are even more promising.
Biomedicine
Applying the knowledge gained from developed network models and using models described in the literature in a practical case is crucial for effectively transferring knowledge to the real world. Therefore, together with my second thesis supervisor, PhD Jose Manuel Jerez Aragones, and in collaboration with Drs. Emilio Alba and Nuria Ribelles, we have been actively dedicated to an application in oncological clinical information management, Galen. Our goal is to implement these models in real data, advancing various lines of practical research that directly address the needs and challenges of oncology.
Managing clinical information in oncology presents unique challenges due to the complexity and sensitivity of the involved data. We are focused on developing methodologies that not only improve operational efficiency in data management but also optimize clinical decision-making. This includes applying advanced data mining techniques and machine learning to identify significant patterns in large clinical datasets, which is crucial for improving diagnostic accuracy and patient prognosis.
One of our research lines focuses on integrating multimodal data such as medical images, genetic data, and treatment records to provide a comprehensive view of each patient's condition. This not only facilitates personalized care but also enables the identification of predictive biomarkers that can influence therapeutic decisions. Additionally, we are exploring the application of predictive models based on neural networks and deep learning techniques to predict individual responses to specific treatments, paving the way towards personalized medicine in oncology.
In our approach, we not only consider the effectiveness of theoretical models but also their feasibility and practical applicability in real clinical settings. This involves close collaborations with medical teams and oncology specialists to ensure that our models are not only technically accurate but also practical and useful in daily practice. We have established a continuous feedback cycle where the results of our models are constantly validated with real clinical data, allowing us to refine and improve our methodologies iteratively.
Furthermore, our work extends beyond the development of algorithms and computational models. We are also dedicated to educating and training healthcare professionals and clinical researchers in the use of these technologies. We strongly believe in the importance of empowering the next generation of doctors and scientists to harness the potential of artificial intelligence and machine learning for the benefit of patients and the broader medical community.
Our multidisciplinary collaboration with experts in computer science, bioinformatics, and oncology has enabled us to advance the frontier of knowledge in this field. We are committed to publishing our findings in high-impact scientific journals and presenting our results at international conferences to share our knowledge and foster the exchange of ideas within the global scientific community.
In summary, our work in applying neural network models and advanced machine learning techniques to oncological clinical information management aims not only to advance scientific knowledge but also to positively transform clinical practice. We are dedicated to developing innovative solutions that can have a significant impact on the diagnosis, treatment, and monitoring of cancer patients, paving the way towards a more personalized, precise, and effective medicine.
