PhD thesis by Sandra Diaz Pier
IAS Series Vol. 47:
Structural plasticity as a connectivity generation and optimization algorithm in neural networks
Sandra Diaz Pier
ISBN 978-3-95806-577-2, https://hdl.handle.net/2128/28830
Schriften des Forschungszentrums Jülich, IAS Series 47, 2021, 167 pages
Structural plasticity as a connectivity generation and optimization algorithm in neural networks (pdf, 16 MB)
Our brains are formed by networks of neurons and other cells which receive, filter, store and process information and produce actions. The morphology of the neurons changes through time as well as the connections between them. For years the brain has been studied as a snapshot in time, but today we know that the way it structurally changes is strongly involved in learning, healing, and adaptation. The ensemble of structural changes that neural networks present through time is called structural plasticity. In this work, I present structural plasticity from its neurobiological foundations and the implementation of a model to describe generation and optimization of connectivity in spiking neural networks. I have targeted two relevant and open questions in the computational neuroscience community: how can we model biologically inspired structural changes in simulations of spiking neural networks and how can we use this model and its implementation to optimize brain connectivity to answer specific scientific questions related to healing, development, and learning. I present several studies which explain the implementation of structural plasticity in a well established neural network simulator and its application on different types of neural networks. In this thesis I have also defined the requirements and use cases for the co-development of tools to visualize and interact with the structural plasticity algorithm. Moreover, I present two scientific applications of the structural plasticity model in the clinical neuroscience and computer science fields. In conclusion, my thesis provides the basis of a software framework and a methodology to address complex neuroscience questions related to plasticity and the links between structure and function in the brain, with potential applications not only in neuroscience but also for machine learning and optimization.