Summary: Researchers have developed a new computer system that emulates the brain’s neural networks. This advance may improve understanding of neural processing and support research into brain disorders.
Source: Frontiers
Researchers report that a custom-built computer designed to mimic brain networks produces results comparable to leading brain-simulation supercomputer software. The system, called SpiNNaker, was evaluated for accuracy, runtime and energy efficiency against NEST, a specialist neural-simulation software, and shows promise for large-scale, low-power simulations of cortical networks. Such capabilities could accelerate studies of learning, robotics and neurological disorders including epilepsy and Alzheimer’s disease.
SpiNNaker is a neuromorphic computing platform developed as part of the Human Brain Project’s Neuromorphic Computing Platform. Over roughly 15 years, engineers and neuroscientists designed SpiNNaker to reflect key features of the brain’s structure and function. The hardware contains hundreds of thousands of simple processors distributed across multiple circuit boards and runs its own software stack optimized for event-driven neural simulation.
According to Dr. Sacha van Albada, lead author and head of the Theoretical Neuroanatomy group at the Jülich Research Centre, “SpiNNaker supports detailed biological models of the cortex—the brain’s outer layer responsible for processing sensory input—and produces outcomes closely matching those obtained with equivalent supercomputer software simulations. Running large-scale, biologically detailed networks in real time with low power consumption will drive progress in robotics and enable more extensive studies of learning and brain disorders.”
The human brain contains roughly 100 billion neurons connected through trillions of synapses. Neuroscience has advanced our understanding of single-neuron dynamics and the functions of many brain regions, but linking neural activity to behavior—such as converting thought into coordinated movement—remains a major challenge. Simulating large neural networks helps bridge that gap, yet conventional supercomputers face limits in speed and energy consumption. Even the best software on the fastest clusters can simulate only a small fraction of the human brain in real time.
Professor Markus Diesmann, co-author and head of Computational and Systems Neuroscience at the Jülich Research Centre, emphasizes the technical barriers: “It remains unclear which computer architectures are optimal for efficient whole-brain simulations. Current high-performance systems often take minutes to simulate one second of biological time, making long-term processes like learning—occurring over hours or days—difficult to study. The brain is also far more energy efficient than standard supercomputers. Neuromorphic, or brain-inspired, computing explores how close electronic systems can come to that efficiency.”
In the study, the team compared SpiNNaker’s performance with NEST, a widely used neural-network simulation package running on a high-performance cluster. The researchers ran a full-scale cortical microcircuit model containing approximately 80,000 neurons and around 0.3 billion synapses—one of the largest models simulated on SpiNNaker to date. This microcircuit is representative of local cortical structure and provides a useful benchmark for scaling toward larger cortical simulations.
Steve Furber, co-author and Professor of Computer Engineering at the University of Manchester, notes, “Simulations executed on NEST and SpiNNaker produced very similar results. This is the first time such a detailed cortical simulation has been run on SpiNNaker or any neuromorphic platform. SpiNNaker’s hardware comprises roughly 600 circuit boards and over 500,000 processors; the experiment in this study used six boards, representing just 1% of the full machine. Our software improvements aim to reduce that requirement to a single board for comparable simulations.”
The evaluation examined numerical accuracy, real-time performance and energy use. SpiNNaker uses fixed-point arithmetic and event-driven processing, while NEST typically employs floating-point arithmetic on conventional cluster hardware with hybrid parallelization. To match the accuracy of NEST running with 0.1 ms integration steps, SpiNNaker required a slowdown factor on the order of 20 relative to real time for this cortical model. NEST’s runtime can reach about three times real time under hybrid parallelization, but that faster runtime correlates with higher power and energy consumption. The study found operating points where energy per synaptic event became comparable between the two platforms, highlighting trade-offs between speed and energy.
Van Albada outlines future directions: “We aim to scale up neuromorphic simulations toward increasingly large real-time networks. Within the Human Brain Project we are collaborating with neurorobotics groups who intend to use neuromorphic systems like SpiNNaker for robot control and embodied cognition experiments.”
Funding: European Union; European Union Seventh Framework Program; UK Engineering and Physical Sciences Research Council; European Union’s Horizon 2020 Research and Innovation Program; European Research Council; Michael Bontenackels.
Source: Emma Duncan – Frontiers
Publisher: Organized by NeuroscienceNews.com
Image Source: Public domain image from NeuroscienceNews.com
Original Research: Performance Comparison of the Digital Neuromorphic Hardware SpiNNaker and the Neural Network Simulation Software NEST for a Full-Scale Cortical Microcircuit Model by Sacha J. van Albada, Andrew G. Rowley, Johanna Senk, Michael Hopkins, Maximilian Schmidt, Alan B. Stokes, David R. Lester, Markus Diesmann and Steve B. Furber. Frontiers in Neuroscience. Published May 23, 2018. doi: 10.3389/fnins.2018.00291
Abstract
Performance Comparison of the Digital Neuromorphic Hardware SpiNNaker and the Neural Network Simulation Software NEST for a Full-Scale Cortical Microcircuit Model
The SpiNNaker platform was developed to enable large-scale neural network simulations in real time while minimizing power consumption. Real-time performance is based on 1 ms integration time steps, appropriate for networks where faster dynamics can be neglected. By slowing down simulations, shorter integration time steps and faster biologically relevant dynamics can be included. This work reports one of the first full-scale cortical microcircuit simulations on SpiNNaker using biological time scales. The model—about 80,000 neurons and 0.3 billion synapses—is the largest run on SpiNNaker so far. Recent software advances permit spreading simulations across boards, facilitating scale-up to larger cortical circuits. Comparison with NEST on a high-performance cluster shows that both platforms can achieve similar accuracy despite SpiNNaker’s fixed-point arithmetic. Runtime and power consumption were assessed: to match NEST’s accuracy with 0.1 ms time steps, SpiNNaker required an approximately 20-fold slowdown relative to real time. NEST runtime can plateau near three times real time when using hybrid MPI and multi-threading, but this comes with increased energy use. The lowest total energy consumption for NEST was observed at around 144 parallel threads and a 4.6-fold slowdown, where energy per synaptic event was comparable to SpiNNaker. These findings broaden SpiNNaker’s application domain and guide further optimizations—such as synapse-centric network representations—needed to enable real-time simulation of large biological neural networks.