By Steve Furber
Following on the footsteps of global investments in brain research, including the United States’ BRAIN Initiative and the European Union’s Human Brain Project (HBP), major efforts are under way to decipher the brain’s complex functions through computational models. In the United Kingdom, research such as the SpiNNaker project—a novel computer architecture inspired by the working of the human brain—may hold the key to new understanding of the ways in which the brain works as well as new insights into neurological diseases. Steve Furber, ICL Professor of Computer Engineering in the School of Computer Science at the University of Manchester, UK, is involved with both SpiNNaker and the HBP.
Figure 1: A plot of the SpiNNaker chip.
The SpiNNaker project team is currently building a massively-parallel chip multiprocessor system for modeling large systems of spiking neurons in real-time. The largest SpiNNaker machine will be capable of simulating a billion simple neurons, or millions of neurons with complex structure and internal dynamics. This work entails both using the computer power now available to improve our understanding of the brain as well as using knowledge of the brain to build better computers.
SpiNNaker will provide a research tool for neuroscientists, computer scientists, and roboticists as it furthers the investigation of new computer architectures that break the rules of conventional supercomputing. In tandem, the University of Heidelberg, also a partner in HBP, is conducting work along similar lines with brain-inspired multiscale computation in neuromorphic hybrid systems (BrainScaleS). BrainScaleS aims to understand and emulate the function and interaction of multiple spatial and temporal scales in brain information processing using both in-vivo experimentation and computational analysis.
Figure 2: The 48-node SpiNNaker board incorporating 864 ARM.
The Heidelberg system operates 104 times faster than biological real-time, but it casts the models into the circuit, so decisions have to be made about what models to support at design time. SpiNNaker operates at the same speed as biological real-time and has a very lightweight communications system albeit supporting a lower overall data rate. Each neuron spike is represented as a single data packet flowing through the communication fabric from the source neuron that generated it to many, possibly thousands, of destination neurons. The neural model is expressed in software, so can support a wide range of models and add new models as users require them. This approach is flexible and, as with its synaptic learning models, SpiNNaker is arbitrarily adaptable (Figures 1 and 2). It requires a huge computing resource and currently occupies a 19-inch rack with 100,000 cores, which is roughly the scale and complexity of a mouse brain (Figures 3 and 4)). The plan is to deliver a system to HBP five times that large, though the goal is to scale up ten-fold.
Figure 3: Pictured is a card frame with 25 SpiNNaker boards, equivalent to 20,000 ARM.
A team from Jülich in Germany is already running a microcortical model on SpiNNaker, having previously used a supercomputer, and a team from Stockholm is using it to model a Bayesian probability network. The focus is on the science in the hope that we can remove the computational limitations on what most computational neuroscience groups have already done. Neural network research can both abstract away from the biological details to understand the information processing principles at work in the brain, and also go deep into detailed biological models to improve understanding of the biology. HBP is a science-first approach that is identifying models that can then develop an ICT approach to understand the workings of the brain better, with the objective ultimately to lead to better treatments for brain disease and novel brain-inspired ICT architectures. Projects such as SpiNNaker have the potential to take us there.
Figure 4: The SpiNNaker computing resource—a rack with 3 card frames, 100,000 ARM cores, which is roughly equivalent to the scale of a mouse brain.
- S. B. Furber, F. Galluppi, S. Temple, and L. A. Plana, ”The SpiNNaker Project,” IEEE Proceedings, vol. 102, no. 5, pp. 652-665, 2014. ISSN 0018-9219
- S. B. Furber, D. R. Lester, L. A. Plana, J. D. Garside, E. Painkras, S. Temple, and A. D. Brown, “Overview of the SpiNNaker system architecture Computers,” IEEE Transactions on Computers, vol. 62, no. 12, pp. 2454-2467, 2013. ISSN 0018-9340
- E. Painkras, L. A. Plana, J. Garside, S. Temple, F. Galluppi, C. Patterson, D. R. Lester, A. D. Brown, and S. Furber. ”SpiNNaker: A 1W 18-core System-on-Chip for Massively-Parallel Neural Network Simulation,” IEEE Journal of Solid-State Circuits, vol. 48, no. 8, pp. 1943-1953, 2013. ISSN 0018-9200
- S. Furber, “TO BUILD A BRAIN—Getting to the bottom of how our brains work is a monumental task, but some innovative computational tricks and a million ARM processors could help.” IEEE Spectrum, vol. 49, no.8, pp. 44-49, Aug 2012. ISSN 0018-9235
Steve Furber CBE FRS FREng is Imperial College London (ICL) Professor of Computer Engineering in the School of Computer Science at the University of Manchester, UK. There, he leads research into asynchronous and low-power systems and, more recently, neural systems engineering.