Assembly embeddings and their neuronal signatures

• Applicants: Professorin Dr. Sonja Grün; Professor Dr. Moritz Helias; Professor Michael Schaub, Ph.D.
• Subject Area: Experimental and Theoretical Network Neuroscience ; Bioinformatics and Theoretical Biology
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 561027837

The brain differs from artificial networks by employing pulsed communication with spikes. Experimental evidence has accumulated that spikes form precisely-timed spatio-temporal patterns that occur in relation to sensory-motor processing and behavior. The current proposal aims to understand whether precisely-timed spikes are essential for robust and efficient information processing. This question is a key challenge in neuroscience and its answer would open the door to novel forms of robust and energy-efficient artificial neural networks.We address this challenge by combining theory and experiment. The theoretical approach studies computation in analog neuromorphic circuits and realistic simulations by contrasting two stereotypical forms of computation: In paradigm 1) networks are trained with time-to-first-spike learning rules that lead to well-timed spikes implementing network function. Paradigm 2) employs reservoir computing, in which spiking neurons tend to fire asynchronously and network dynamics serve to embed stimuli into a high-dimensional space, a subspace of which implements the desired function. These two alternative paradigms have drastically different ramifications for the observable brain activity and graph structure of the network’s connectivity; while paradigm 1) leads to precisely-timed spikes and feed-forward, directed connectivity, paradigm 2) features asynchronous firing and networks that are recurrent and random. Investigating connectivity by higher-order graph signal processing will allow us to identify functional signatures for either hypothesis that subsist even under subsampling. To bridge to brain responses we apply a multi-scale approach and utilize spiking models as the basis for modeling population signals (VSDI, MEG). Predictions for these signals require embedding of networks into a sheet of cortex. Such virtual recordings will expose limitations of current methods to distinguish between the two paradigms and help us develop and sharpen respective analysis methods before applying them to experimental data. Detection of precise spike timing within parallel spike recordings is challenging due to massive sub-sampling. We therefore obtain two new datasets, in addition to spike recordings, which trade spatial resolution for a more complete sampling and larger spatial extent. 1) Voltage-sensitive dyes imaging (VSDI) in monkeys will provide millisecond time resolution while spanning multiple areas of the visual system at a resolution as high as a few 10s of cells per pixel. VSDI depends on the neurons’ summed inputs and is hence highly sensitive to synchronous appearance of spikes. 2) Full brain MEG recordings in humans allow us to reach the maximum possible spatial scale, still at high temporal resolution. The goal is to distinguish between the two paradigms by neuronal signature derived from iterative adjustments of models validated on experimental data.

• DFG Programme: Research Grants
• International Connection: Israel, Switzerland
• Partner Organisation: The Israel Science Foundation
• Cooperation Partners: Dr. Mihai A. Petrovici; Professorin Dr. Hamutal Slovin; Professorin Dr. Mina Teicher