Talks

Title: Large-scale network models as digital twins advance theory and neuromorphic computing

Abstract: Computational neuroscience is entering a new era. This originates from the convergence of two developments: First, knowledge has been accumulated enabling the construction of anatomically detailed models of one or multiple brain areas. The models have cellular and synaptic resolution, represent the respective part of the brain with its natural number of neurons and synapses, and are multi-scale. Next to spiking activity, also mesoscopic signals like the local field potential (LFP) and fMRI signals can be generated (e.g. [1]). Second, with the completion of the European Human Brain Project (HBP), simulation has firmly established itself in neuroscience as a third pillar alongside experiment and theory. A conceptual separation has been achieved between concrete network models and generic simulation engines [2,3]. Many different models can be simulated with the same engine, such that these simulation codes can continuously be optimized and operated as an infrastructure [4]. Network models with millions of neurons can routinely be investigated.

Neuroscientists can now work with digital twins of certain brain structures to test their ideas on brain functions and probe the validity of approximations required for analytical approaches. However, the efficient use of this new capability also requires a change in mindset.

Computational neuroscience seems stuck at a certain level of model complexity for the last decade not only because anatomical data were missing or because of a lack of simulation technology. The fascination of the field with minimal models leads to explanations for individual mechanisms, but the reduction to the bare equations required provides researchers with few contact points to build on these works and construct larger systems with a wider explanatory scope. In addition, constructing large-scale models goes beyond the period of an individual PhD project, but an exclusive focus on hypothesis-driven research may prevent such sustained constructive work. Possibly, researchers may also just be missing the digital workflows to reuse large-scale models and extend them reproducibly. The change of perspective required is to view digital twins as research platforms and scientific software as infrastructure with all consequences for the requirements on quality, long-term availability, and support.

As a concrete example, the presentation discusses how the universality of mammalian cortex has acted as a motivation to construct large-scale models and demonstrates how digital workflows have helped to reproduce results and increase the confidence in such models.

A digital twin promotes neuroscientific investigations, but can also serve as a benchmark for technology. The talk shows how a model of the cortical microcircuit has become a de facto standard for neuromorphic computing [5].

[1] Senk J., Hagen E, van Albada SJ, Diesmann M (2018) Reconciliation of weak pairwise spike-train correlations and highly coherent local field potentials across space. arXiv:1805.10235 [q-bio.NC]

[2] Einevoll GT, Destexhe A, Diesmann M, Grün S, Jirsa V, de Kamps M, Migliore M, Ness TV, Plesser HE, Schürmann F (2019) The Scientific Case for Brain Simulations. Neuron 102:735-744

[3] Senk J., Kriener B., Djurfeldt M., Voges N., Jiang HJ., Schüttler L., Gramelsberger G., Diesmann M., Plesser HE., van Albada SJ. (2022) Connectivity concepts in neuronal network modeling. PLOS Comput Biol 18(9):e1010086

[4] Aimone JB, Awile O, Diesmann M, Knight JC, Nowotny T, Schürmann F (2023) Editorial: Neuroscience, Computing, Performance, and Benchmarks: Why It Matters to Neuroscience How Fast We Can Compute. Front Neuroinform 17. DOI:10.3389/fninf.2023.1157418

[5] Kurth AC., Senk J., Terhorst D., Finnerty J., Diesmann M. (2022) Sub-realtime simulation of a neuronal network of natural density. Neuromorphic Computing and Engineering 2:021001

Title: Nightmares of spike pattern analysis

Abstract: It is believed that neuronal interaction in the cortex is organized in cell assemblies. The signature of such cell assembly activities is hypothesized as the occurrence of spatio-temporal spike patterns (STPs). The analysis of massively parallel spike data for such STPs requests smart algorithms and compute power. The way of identifying statistical significance of such repeating STPs and the proof that they are specific to behavior provides many obstacles. I will report here on some of them, in particular the occurrence of artifacts and the problem of removing them.

Title: Reconstructing neuronal circuitry from spiking signals from multiple neurons

Abstract: State-of-the-art techniques allow researchers to record large numbers of spike trains in parallel for many hours. Such recordings will enable us to infer the fine structure of neural circuits, i.e., the synaptic connectivity between neurons, which clarifies how the brain computes based on neurons. Cross-Correlation (CC) method is a standard method for estimating the connectivity from parallel spike trains (Perkel et al., 1967). While the CC method has been used to estimate neuronal connectivity, its estimate becomes unreliable when neural activity fluctuates largely. In this presentation, we propose two approaches to resolve the large fluctuation problem: Generalized Linear model for Cross-Correlation (GLMCC) (Kobayashi et al., 2019) and COnvolutional Neural Network for Estimating synaptic ConnecTivity (CONNECT) (Endo, Kobayashi, et al., 2021). We demonstrate that these methods can robustly estimate the synaptic connectivity from parallel spike trains by applying them to two synthetic datasets generated by the network of spiking neurons and showed that the proposed methods performed better than conventional methods (e.g., jittering method: Amarasingham et al., 2012). A ready-to-use version of the web application, the source code, and example data sets are available on our website (https://s-shinomoto.com/CONNECT).

Title: Instantaneous firing rate and counting statistics of spike trains

Title: Encoding of fluctuating and constant pheromone signal by moth olfactory receptor neurons

Title: Non-differentiable activity in the brain

Abstract: Spike rasters of multiple neurons show vertical stripes, indicating that neurons exhibit synchronous activity in the brain. We seek to determine whether these coherent dynamics are caused by smooth brainwave activity or by something else. By analyzing biological data, we find that their cross-correlograms exhibit not only slow undulation but also a cusp at the origin, in addition to possible signs of monosynaptic connectivity. Here we show that undulation emerges if neurons are subject to smooth brainwave oscillations while a cusp results from non-differentiable fluctuations. While modern analysis methods have achieved good connectivity estimation by adapting the models to slow undulation, they still make false inferences due to the cusp. We devise a new analysis method that may solve both problems. We also demonstrate that oscillations and non-differentiable fluctuations may emerge in simulations of large-scale neural networks.

Title: Realistic Modeling of Spike Train data through Point Processes

Abstract: Point process theory has been extensively explored to model electrophysiological data. The classical approach consists in using established models and assuming stationarity of the neuronal firing rate. However, whenever non-stationary processes are generated, the proposed model rarely includes further statistical features of experimental data, such as regularity, dead time and higher-order correlations.

In this talk, I will introduce the statistical features that need to be taken into consideration in order to closely model an exemplary recording session of electrophysiological data. The statistics include non-stationary firing rate, dead time, regularity, pairwise and higher-order correlations. Furthermore, I will review the existing point process models, techniques and tools to generate artificial data with such statistical features. First, I will introduce how surrogate techniques can be used to test particular properties of spike trains. I will take into consideration 6 different techniques and examine their statistical properties such as spike loss, ISI characteristics, effective movement of spikes, and arising false positives when applied to different ground truth data sets: first, on stationary, and then on non-stationary point processes models mimicking statistical properties of experimental data.

Then, I will introduce five artificial data sets, all modeling some of the statistical characteristics of a experimental recording session of cortical data of a macaque monkey. I analyze all simulated data sets with different techniques and compare their statistics to the original data. The generated data sets reproduce the statistical complexity of experimental data with increasing degree, while being fully artificial and generated in a controlled way. Thus, they can be employed as ground truth data for testing and benchmarking of existing and future methods for the analysis of parallel spike trains. Finallt, they can be used for didactic purposes, in order to approach experimental data in the early stages of study and research.

Title: Network inference in a stochastic multi-population neural mass model via approximate Bayesian computation