My PhD thesis focused on studying the information-theoretic, statistical, and physical aspects of machine learning algorithms. My research is an attempt to answer two fundamental questions in this field: 1) How to measure the learned information in training neural networks? 2) Where this learned information stored? During the course of my PhD, I have developed a quantitative method, in response to the first question, that make it possible to measure the learned information inside neural networks. This method leverage the toolbox of stochastic thermodynamics that connect information-theoretic concepts and physics laws. Furthermore, this method shows the flow of learned information into parameter space that is the first clue in answering the second question. Finally, the localization of the learned information in parameter space has been achieved by redefining the role of hidden neurons in neural networks.
Prior to my PhD, I did two years of research in Computational Neuroscience during my M.S. degree education. My master’s thesis centered on studying the role of functional hubs in directing flow of information in neuronal networks. This research evolves mathematical modeling and simulation of neuronal activities in presence of local stimulus.