Regensburg 2025 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
SOE: Fachverband Physik sozio-ökonomischer Systeme
SOE 7: Focus Session: Self-Regulating and Learning Systems: from Neural to Social Networks
SOE 7.8: Vortrag
Mittwoch, 19. März 2025, 12:00–12:15, H45
Transient Recurrent Dynamics Shapes Representations in Mice — •Lars Schutzeichel1,2,3, Jan Bauer1,4, Peter Bouss1,2, Simon Musall3, David Dahmen1, and Moritz Helias1,2 — 1Institute for Advanced Simulation (IAS-6), Jülich Research Centre, Germany — 2Department of Physics, Faculty 1, RWTH Aachen University, Germany — 3Institute of Biological Information Processing (IBI-3), Jülich Research Centre, Germany — 4The Edmond and Lily Safra Center for Brain Sciences, The Hebrew University of Jerusalem, Israel
Different stimuli evoke transient neural responses, but how is stimulus information represented and reshaped by local recurrent circuits? We address this question using Neuropixels recordings from awake mice and recurrent network models, inferring stimulus classes (e.g., visual or tactile) from activity. A two-replica mean-field theory reduces complex network dynamics to three key quantities: the mean population activity (R) and overlaps (Q=, Q≠), reflecting response variability within and across stimulus classes. The theory predicts the time evolution of R, Q=, and Q≠. Validated in experiments, it reveals how inhibitory balancing governs the dynamics of R, while chaotic dynamics shape overlaps, providing insights into the mechanisms underlying transient stimulus separation. The analysis of mutual information of an optimally trained population activity readout reveals that sparse coding (small R) allows the optimal information representation of multiple stimuli.
Keywords: recurrent neural networks; population coding; mean-field methods; information theory; Spin Glass diagrammatics