Desarrollo de nuevos algoritmos neurogliales que modelizan la interacción astrocito-neurona en sistemas de altas prestaciones

Abstract

We are living an era of constant evolution due to technological advances. Many of them are being possible thanks to Artificial Intelligence (AI) and the large volumes of data are being stored. You can build tools with intelligent components that are revolutionizing various fields thanks to a large amount of data that can be analyzed by these intelligent models. However, many investigations only focus on the quantity and quality of the data available, and little effort is made to improve the AI techniques themselves. The proposal of this thesis is the improvement of connectionist intelligent systems that until recently were formed by Artificial Neural Networks (ANN). To carry out such ambitious work, it has been considered, as has happened with research in other areas, to consider how nature solves the problem. For this, we will focus on the most complex and efficient structure known, the human brain. To that end, it is necessary to rely on the field of Neuroscience where one can try to take the advances that are discovered or the hypotheses that are generated in the AI field. The core of this thesis is based on research that shows that neurons are not the only elements of the human brain that participate in the processing of information. Astrocytes of the glial system play an essential role in the treatment of information. In fact, it is known that synaptic communication occurs with the participation of neurons and astrocytes, which is known as tripartite synapses. This led to the inclusion of new elements that simulate the behaviour of glial cells in ANN. The addition of the new elements, artificial astrocytes, originated the Artificial NeuroGial Networks (ANGN). To demonstrate the usefulness of artificial astrocytes and collaborate in demonstrating the capacity of the Glial System (GS), new algorithms of astrocytic modulation have been used. These algorithms have been tested in different classification and regression problems, obtaining significant results with respect to networks that do not use the GS. In addition, an open source web application has been developed thanks to this thesis so that the scientific community can use these networks freely.