Usage and Scaling of an Open-Source Spiking Multi-Area Model of Monkey Cortex

Sacha J. van Albada, Jari Pronold, Alexander van Meegen, Markus Diesmann

Abstract

Researchers have developed a massive, open-source spiking neural network model of the macaque monkey’s visual cortex, featuring 4 million neurons and 24 billion synapses. By bridging the gap between microscopic neuron activity and macroscopic brain imaging (fMRI), this project provides a reproducible "executable specification" for large-scale neuroscience and demonstrates how these models scale on modern supercomputers.

Introduction

Computational neuroscience is currently facing a "complexity barrier." While numerous small-scale models exist for specific brain circuits, integrating them into a single, large-scale framework consistent with diverse experimental data—including anatomy, spike rates, and fMRI—remains a significant challenge. Most existing models are either too small to capture whole-brain dynamics or too abstract to represent individual neurons. Furthermore, these massive models are often "black boxes," making them difficult to reproduce or modify without specialized expertise.

This paper addresses these issues by providing a fully documented, open-source workflow that relates the connectivity of 32 vision-related cortical areas to their resting-state dynamics.

Methods and Architecture

The model is constructed using the NEST simulation engine, a standard tool for spiking neural networks. The architecture represents 32 areas of the macaque cortex, where each area is modeled as a 1 $mm^2$ microcircuit with four layers (2/3, 4, 5, and 6). Each layer contains excitatory and inhibitory populations of integrate-and-fire neurons.

Multi-area connectivity architecture and laminar patterns.

To define connectivity, the authors utilized "predictive connectomics" to fill gaps in anatomical data, applying statistical rules based on neuron densities and inter-area distances. Data sources included the CoCoMac database for existence and laminar patterns, quantitative tracing data for source neuron counts, and cortical architecture data for local microcircuit constraints. To ensure biological plausibility, a semi-analytical mean-field method was employed to stabilize firing rates. The entire workflow is managed via Snakemake to ensure reproducibility from raw data processing to final visualization.

Key Insights and Results

A primary finding is that the "resting-state" temporal structure only emerges when the full multi-area model is simulated; simulating only a subset (e.g., V1 and V2) fails to reproduce the slow activity fluctuations observed in vivo. This suggests that the whole is indeed greater than the sum of its parts in cortical dynamics.

Spiking activity across cortical layers in area V1.

The model represents a significant leap in scale, handling approximately 4 million neurons and 24 billion synapses. This scale requires a substantial memory footprint, necessitating roughly 1 TB of RAM. On 160 compute nodes of the JURECA supercomputer, the model achieves a real-time factor of 31 (1 second of biological time per 31 seconds of wall-clock time).

The study also highlights computational bottlenecks: network construction (instantiating 24 billion synapses) is a major hurdle, often dominating total runtime. While increasing supercomputer nodes (strong scaling) reduces neuron update time, MPI communication overhead eventually becomes the limiting factor as spike data must be shared across more nodes.

Discussion and Future Directions

While the model provides a powerful framework, several areas for improvement remain. The authors note that network construction still incurs significant Python overhead, suggesting that moving setup logic into C++ could enhance performance. Additionally, the current "round-robin" distribution of neurons across nodes is sub-optimal; future iterations might distribute neurons based on the brain's modular structure to minimize inter-node traffic.

Currently, the model uses static synapses. Incorporating plasticity and learning is a major future goal, though it will require even greater computational resources. Finally, the authors advocate for tighter integration between code repositories and high-performance computing infrastructure to make these "executable specifications" more accessible to the wider scientific community.

Summary updated

Abstract

We are entering an age of ‘big’ computational neuroscience, in which neural network models are increasing in size and in numbers of underlying data sets. Consolidating the zoo of models into large-scale models simultaneously consistent with a wide range of data is only possible through the effort of large teams, which can be spread across multiple research institutions. To ensure that computational neuroscientists can build on each other’s work, it is important to make models publicly available as well-documented code. This chapter describes such an open-source model, which relates the connectivity structure of all vision-related cortical areas of the macaque monkey with their resting-state dynamics. We give a brief overview of how to use the executable model specification, which employs NEST as simulation engine, and show its runtime scaling. The solutions found serve as an example for organizing the workflow of future models from the raw experimental data to the visualization of the results, expose the challenges, and give guidance for the construction of an ICT infrastructure for neuroscience.
BibTeX
@inproceedings{albada2019usage,
  author = {Sacha J. van Albada and Jari Pronold and Alexander van Meegen and Markus Diesmann},
  title = {{Usage and Scaling of an Open-Source Spiking Multi-Area Model of Monkey Cortex}},
  booktitle = {BrainComp},
  year = {2019},
  doi = {10.1007/978-3-030-82427-3\_4},
}
Figure