PhD thesis supervisor: dr. Kęstutis Pyragas (apply for recommendation)
Next-generation reservoir computing for neural network dynamics prediction
The aim of the proposed PhD topic is to develop and advance modern machine learning algorithms for modeling and predicting the macroscopic behavior of complex neuronal networks. Real neuronal networks are highly complex dynamical systems composed of billions of interacting neurons. In neuroscience, understanding and predicting the macroscopic behavior of such systems is crucial, as it underlies essential physiological functions of neural networks, and their dysfunctions may lead to neurological diseases.
Direct modeling of neuronal networks at the microscopic level faces extremely large computational challenges. To overcome this complexity, phenomenological models have been developed that mimic the behavior of large neuronal populations at the macroscopic level (so-called macroscopic models). However, such models often exhibit limited accuracy. Recently, a method has been proposed that allows exact macroscopic models to be derived directly from microscopic neuronal equations; however, this method is not universal and is applicable only to very simple neuron models.
This PhD project proposes the use of machine learning methods to automatically learn macroscopic models of neuronal networks from microscopic models’ dynamical data. The focus will be on algorithms based on next-generation reservoir computing (NGRC), which does not require an explicit reservoir. During the PhD studies, the NGRC method will be further developed and applied to the construction of macroscopic models of complex neuronal networks and to the prediction of their macroscopic dynamics.