P274 Dynamics of sensory stimulus representations in recurrent neural networks and in mice
Lars Schutzeichel*1,2,3, Jan Bauer1,4,5, Peter Bouss1,2, Simon Musall3, David Dahmen1and 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 4Gatsby Unit for Computational Neuroscience, University College London, United Kingdom 5Edmond and Lily Safra Center for Brain Sciences, The Hebrew University of Jerusalem, Israel
The information about external stimuli is encoded in the responses of neuronal populations in the brain [1,2], forming neural representations of the stimuli. The diversity of responses is reflected in the extent of these neural representations in neural state space (Fig. 1a). In recent years, understanding the manifold structure underlying neuronal responses [3] has led to insights into representations in both artificial [4] and biological networks [5]. Here, we extend this theory by examining the role of recurrent network dynamics in deforming stimulus representations over time and their influence on stimulus separability (Fig. 1b). Furthermore, we assess the information conveyed for multiple stimuli (Fig. 1c). Methods We simulate recurrent networks of binary neurons and study their dynamics analytically using a two-replica mean-field theory, reducing the dynamics of complex networks to only three relevant dynamical quantities: the population rate and the representation overlaps within and between stimulus classes. These networks are fit to Neuropixels recordings from the superior colliculus of awake behaving mice. To assess the information conveyed by multiple stimuli, we analyze the mutual information between an optimally trained readout and the stimulus class. To calculate the overlap of representations within and across stimulus classes, we utilize spin glass methods [6]. Results Stimulus separability and its temporal dynamics are shaped by the interplay of three dynamical quantities: the mean population activity E and the overlaps θ= and θ≠, which represent response variability within and across stimulus classes, respectively (Fig. 1b). For multiple stimuli, there is a trade-off: as the number of stimuli increases, more information is conveyed, but stimuli become less separable due to their growing overlap in the finite-dimensional neuronal space (Fig. 1c). We find that the experimentally observed small population activity R lies in a regime where information grows extensively with the number of stimuli, sharply separated from a second regime in which information converges to zero. Discussion Separability is a minimal requirement for meaningful information processing: The signal propagates to downstream areas, where, along the processing hierarchy, representations of different perceptual objects must become increasingly separable to enable high-level cognition. Our theory reveals that sparse coding not only provides a crucial advantage for information representation but is also a necessary condition for non-vanishing asymptotic information transfer. Our work thus provides a novel understanding of how collective network dynamics shape stimulus separability.
Figure 1. Overview. a: Stimulus representations characterized by their distance from the origin R and their extent θ=. b: Temporal evolution of representations of stimuli from two classes. A linear readout quantifies the separability between the classes of stimuli for every point in time. c: The separability measure also determines the information content in the population signal for P≥2 stimuli. Acknowledgements This work has been supported by DFG project 533396241/SPP2205 References [1]https://doi.org/10.1126/science.3749885 [2]https://doi.org/10.1016/j.tics.2013.06.007 [3]https://doi.org/10.1103/PhysRevX.8.031003 [4]https://doi.org/10.1038/s41467-020-14578-5 [5]https://doi.org/10.1016/j.cell.2020.09.031 [6]https://doi.org/10.1088/1751-8121/aad52e
