Doctoral Research Seminar on November 27, 2019


Doctoral Research Seminar

Jonas Braun (CNE – Computational Neuroengineering, TUM): "Timescales of predictability as a tool to analyse neuronal dynamics in V1 and CA1" and Michael Deistler (CNE – Computational Neuroengineering, TUM): "Analysing perturbations in neuroscience models using simulation-based inference"

Abstract: Neuronal populations in the brain are often described as dynamical systems. A fundamental challenge in characterizing neuronal dynamics in vivo is determining the extent to which the empirically observed dynamics in any brain region are intrinsically generated, or extrinsically driven. Extrinsic drivers could be different sensory inputs, behaviour or neural activity in other regions of the brain. A dominance of intrinsic dynamics would mean that future neuronal activity largely depends on current neuronal activity within the region, and cannot be predicted from external factors. We tackle the question of what drives the dynamics of neuronal activity by characterising what is predictive of neuronal activity at different time points in the future. In particular, we analyse neuronal activity in the primary visual cortex (V1) and the hippocampus (CA1) of mice while they either perform a spatial navigation task or passively view different visual stimuli. We find that, in both data sets, extrinsic drivers dominate the dynamics in either area. More specifically, behaviour and sensory input, but not the respective other area, provide the most predictive features for population activity in V1 and CA1. Our analysis of predictability on different time lags offers a way to characterise neuronal dynamics on different levels of the processing hierarchy and extends analyses that have identified strong instantaneous sensory and behavioural encodings.

Michael Deistler (CNE – Computational Neuroengineering, TUM): "Analysing perturbations in neuroscience models using simulation-based inference"

Abstract: Many dynamical systems in biology exhibit sloppiness: their behavior is invariant to some perturbations of parameters, but highly sensitive to others. A classical example in neuroscience is the crustacean stomatogastric ganglion, where it has been shown that similar activity can arise from highly different sets of membrane and synaptic conductances. Here, we present statistical tools for studying how perturbations in parameter space affect the output in dynamical models of biological systems. We approach the investigation of sloppiness as a Bayesian inference problem and demonstrate this approach on a model of the pyloric network of the crustacean stomatogastric ganglion. We show that parameter sets producing similar outputs are connected in parameter space and find that network output can be preserved even under large perturbations. We then investigate how finely parameters have to be tuned to retain similar network output. Our methods allow us to find directions in the parameter space where even small perturbations lead the circuit to break down, demonstrating high sensitivity of network activity to changes in circuit parameters. Lastly, we use the knowledge of the joint distribution over circuit parameters to identify compensation mechanisms, and show that the predicted mechanisms are consistent with experimentally measured data. The simulation-based inference tools presented here will be applicable to study sensitivities and invariances in a wide range of dynamical models of neural circuits.

27 November 2019, 16:30 h Place: ICS Karlstraße 45, 2. Floor, room 2026