Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
SOE: Fachverband Physik sozio-ökonomischer Systeme
SOE 4: Poster Session
SOE 4.8: Poster
Montag, 9. März 2026, 18:00–21:00, P4
Inference of network structures from partial observations — •Matthias Klaus1,2, David Dahmen1, and Moritz Helias1,2 — 1Institute for Advanced Simulation (IAS-6), Jülich Research Centre, Jülich, Germany — 2Department of Physics, Faculty 1, RWTH Aachen University, Aachen, Germany
Today’s recording techniques for neural activity allow one to access finite time windows of a few hundred to thousands of neurons. Since they are embedded in strongly connected networks of 105 neurons, the recorded data will be considerably influenced by the structure of these unobserved parts. Theoretical descriptions of the problem often assume spatial and temporal homogeneity across the system, i.e. that structure and activity of the entire network, on a statistical level, are close to the recorded part.
Here we model this statistical homogeneity by the variances of connectivity and of a globally independent driving white noise. Estimates for these parameters exist from prior experiments. Consequently, we use a Bayesian approach to obtain a posterior for the local connectivity by conditioning on the locally recorded activity. The theory involves a marginalization over unobserved activity, which is exact for a linear network with Gaussian activity. This condition can be relaxed by reasonably assuming a large network and applying dynamic mean-field theory (Sompolinsky, Crisanti, Sommers 1988; Schuecker, Goedeke, Helias 2018). We find that reconstruction of local connectivity is possible if the influence from unobserved parts reduces to a colored noise whose statistics is estimated correctly.
Keywords: Bayesian inference; dynamic mean field theory; recurrent neural network; subsampling; student-teacher model