Help Privacy Policy Disclaimer
  Advanced SearchBrowse





Inferring dynamical states in complex recurrent networks


Buendia,  V       
Department of Computational Neuroscience, Max Planck Institute for Biological Cybernetics, Max Planck Society;


Levina,  A       
Department of Computational Neuroscience, Max Planck Institute for Biological Cybernetics, Max Planck Society;

Fulltext (restricted access)
There are currently no full texts shared for your IP range.
Fulltext (public)
There are no public fulltexts stored in PuRe
Supplementary Material (public)
There is no public supplementary material available

Buendia, V., & Levina, A. (2023). Inferring dynamical states in complex recurrent networks. Poster presented at Bernstein Conference 2023, Berlin, Germany.

Cite as: https://hdl.handle.net/21.11116/0000-000D-D73D-D
Brain activity displays a wide range of behaviours, including asynchronous irregular activity [1], up-and-down states or collective oscillations [2], among others. One relevant open question is to understand the brain's dynamical properties at a large scale from experimental observations coming from electroencephalography (EEG), functional magnetic resonance imaging (fMRI) or magnetic encephalography (MEG). Apart from the functional activity, the network of physiological connections -or connectome- between brain regions can be determined using modern magnetic resonance techniques, such as diffusion tensor imaging (DTI). Understanding how such diversity of dynamical behaviours emerges from the physiological structure would be a significant step towards the comprehension of brain function. However, the heterogeneity among regions and the recurrent structure of the connectome complicates the task of inferring model parameters for each region. Previous approaches use global parameters for the network or simplified models. Here, we introduce a novel method based on simulation-based inference (SBI) to infer the dynamical regimes of regions in a strongly connected network. In recent years, SBI methods have benefitted from technologies such as deep learning, and provide accurate Bayesian inference in complex dynamical systems. At this moment, robust, ready-to-use machine learning technology for scientific contexts [3] is available. Our protocol builds upon these technologies. We study different factors impacting their efficiency, including the low-dimensional representation of the timeseries used to train the network and estimate the parameters, as well as how to sample during the training phase. These two are essential to extract all the information from the series and the phase space of the model that SBI’s deep network learns to later perform the inference. Finally, we test our protocol in synthetic connectome-like networks with a mass model featuring a rich phase diagram, demonstrating that it is possible to obtain estimations of parameters for individual regions, even when in the presence of coupling. One of the main advantages of SBI is that once trained, the network is able to perform inference in a very fast way, allowing it to digest large networks or even entire recording datasets. Thus, preliminary results on fMRI recordings from the Human Connectome Project dataset will be presented.