Such investigations led to the advent in the 1980s of neuromorphic computing, which takes inspiration from the characteristics of neurons and neural networks in the brain. Advances in microchip design and ever-increasing transistor integration density – funded by conventional computing applications but also available to neuromorphic chip designers – helped accelerate development in this field.

"Over the last decade, several large-scale neuromorphic systems have advanced from concept to fruition," explains Steve Furber, ICL Professor of Computer Engineering at the University of Manchester. "These all exploit developments in microchip technology that enable large-scale systems that were previously impractical or unfeasibly expensive."

Furber has recently authored a topical review article examining large-scale neuromorphic computing systems. Writing in the Journal of Neural Engineering, he presents a brief history of the field and takes a look at some of the projects that are now underway (J. Neural Eng. 13 051001).

Furber emphasizes the two-way nature of this research: engineers discover and apply principles from neuroscience to create computing systems, while in return, neuroscientists learn from the results obtained from such systems. He notes that there is a widespread interest in using neuromorphic techniques to model biological processes and enhance understanding of the human brain.

Take four

While neuromorphic systems vary greatly in design, all adhere to the idea that computation is highly distributed across small computing elements analogous in some way to neurons, connected into networks in a similar manner to synapses (which, in the brain, couple the output signal from one neuron into the input of the next). Current neuromorphic capabilities can support neural networks incorporating many millions of neurons with many billions of synapses. Furber describes four large-scale systems currently in development:

• The IBM TrueNorth chip, developed under DARPA's SyNAPSE programme, is based upon distributed digital neural models and aimed at real-time cognitive applications. The key component is a 5.4 million transistor CMOS chip that incorporates 4096 neurosynaptic cores. TrueNorth chips can be connected together to form larger systems, and a 16-chip circuit board has been developed that incorporates 16 million neurons and 4 billion synapses. Designed as an "application delivery" platform to address problems ranging from vision to audition and multi-sensory fusion, TrueNorth offers power-efficient real-time processing of high-dimensional, noisy, sensory data. For example, it can run applications such as real-time object recognition at remarkably low power levels.

• The Stanford Neurogrid chip uses sub-threshold analogue circuits to model neuron and synapse dynamics in biological real time. The Neurogrid system comprises a software suite for configuration and visualization of neural activity, plus hardware (including 16 Neurocore chips and support circuitry) for real-time simulation of the neural network. Each Neurocore chip supports 65,536 sub-threshold analogue neurons and about 100 million synapses. The real-time operation of Neurogrid makes it suitable for robotic control, and it has already been interfaced to a robotic arm with the ultimate goal of controlling a prosthetic limb. Future research will include exploiting the technology's low power to develop a chip that can be implanted in the brain for prosthetic limb control.

• The BrainScaleS system, developed at the University of Heidelberg, uses above-threshold analogue neural circuits. These are much faster than sub-threshold analogue circuits and run at 10,000 times faster than biological real-time. To accommodate this speed-up, BrainScaleS employs wafer-scale integration to deliver large numbers of analogue neurons that can be interconnected efficiently. Within a BrainScaleS wafer, each of 48 reticles holds eight High-Count Analogue Neural Network (HiCANN) die, each of which implements 512 neurons and over 100,000 synapses. BrainScaleS hardware is available in a small portable form with a single HiCANN die, and also as a 20-wafer platform incorporating a substantial cluster server. The high speed of BrainScaleS makes it ideal for applications that take a long time in biological terms, such as long-term learning tasks. Modelling years of childhood development, for example, could potentially be performed within hours.

• The University of Manchester's SpiNNaker system is a massively-parallel digital computer for modelling large-scale spiking neural networks with connectivity similar to the brain, in biological real time. SpiNNaker is based around a custom processing chip containing 18 ARM968 processor cores, with 16,000 neurons and 16 million synapses – a design motivated by the need for scalability and energy-efficiency. SpiNNaker is delivered as a 72-core board for training and small network development, or as an 864-core board that forms the basis of larger systems. Small-scale SpiNNaker systems have been used for applications such as robot control, vision processing and non-real-time modelling of biological circuits. Large-scale applications, though fewer in number, include a real-time implementation of the 2.5 million neuron Spaun cognition model, where SpiNNaker delivers a 9000 times speed-up compared with a high-end desktop machine. Currently, the largest SpiNNaker machine incorporates 500,000 processor cores – the team aims to double this to one million cores over the coming year.

Diverse tactics

Furber notes that the four systems represent a diverse range of approaches to modelling neural systems, with differing objectives. One cannot be considered superior to any other; instead, each has its strengths: TrueNorth enables highly-integrated and energy-efficient application delivery; SpiNNaker offers maximum flexibility for researching neural models and plasticity rules; BrainScaleS brings high acceleration for long-term learning; and Neurogrid offers high energy efficiency with models that are closest to the physics at work in the biology.

"There are many other groups working on neuromorphic technology at smaller scales, some of which are mentioned in the paper," adds Furber. "There are also many interesting developments in neuromorphic sensors and actuators that complement the neural network hardware, and others in purely software systems that support the creation and development of networks that can ultimately benefit from transfer to neuromorphic hardware."

"The field is still very diverse, and over the next few years we can expect some convergence in tools and approaches," Furber told medicalphysicsweb. "Although it may remain a "horses for courses" scenario, where flexible platforms such as SpiNNaker are employed for research and application development and more highly integrated platforms such as TrueNorth are used for application delivery."

Related stories

• Microstimulation could restore lost touch
• Patient's thoughts control robotic arm
• Machine learning diagnoses Parkinson's
• Progress in supercomputing for healthcare