Modeling brain dynamics through recurrent neural networks: a linearization approach

In computational neuroscience, recurrent Neural Networks (RNNs) have recently emerged as a flexible tool to model the dynamics of biological neural circuits. Often, different types of RNNs can adequately model the data, but build on different premises and lead to different descriptions of dynamics. In this thesis, we will review an RNN approach commonly used to model neural dynamics: a reservoir computing approach, where a high-dimensional, linear RNN is trained via linear regression Using both simulated data and real data from functional magnetic resonance imaging, the approach will be systematically analyzed, testing its accuracy in reproducing several types of dynamical systems.