Predicting brain activity from arousal measurements

Several imaging techniques, such as functional magnetic resonance imaging (fMRI) or calcium imaging, allow spatially fine-grained measurements of brain activity in time. By combining an established technique in dynamical systems analysis (time-delayed embedding) with recently developed artificial-neural-network based methods (variational autoencoders), one can retrieve an explicit description of a low-dimensional dynamical system underlying the observed time series. How this low-dimensional dynamics relates to the overall level of arousal, a key physiological parameter, was until recently not known. Adding measurements of arousal (which is non-invasively captured by the pupil diameter) to the autoencoder, a recent publication (Raut et al., 2023) showed that arousal can predict a large part of the observed dynamics. In this thesis, we will review the methodology of Raut et al., apply it to simplified artificial scenarios, and try to reproduce a few results of Raut et al.