Neuromorphic Computers: Promising Candidates for Efficient Brain Simulations

Scientists use simulations of brain activity to investigate learning processes or neurological diseases such as epilepsy and Alzheimer’s disease. At the same time, neural networks are the basis for many applications in artificial intelligence, such as those used to control robots.

The computational effort for such simulations, however, is enormous. At present, only 1 per cent of the human brain can be simulated on the world’s most powerful supercomputers. Moreover, the processes in the simulation run much more slowly than they do in reality. A conventional supercomputer takes several minutes to reproduce one second of biological activity in such a network.

In order to improve the efficiency of the simulation of neuronal networks, scientists are working on so-called neuromorphic computers whose hardware is based on the structure of the human brain. One of the largest and most advanced projects is the SpiNNaker project, which researchers are pursuing as part of the European Human Brain Project.

Scientists at Forschungszentrum Jülich recently carried out the largest simulation of a neural network on a SpiNNaker system to date. They then compared the findings with results they had obtained on a conventional computer. For their study, the researchers calculated the activity in a network of 80,000 neurons connected by more than 300 million synapses. In terms of its size, this corresponds to just one millionth of the human brain.

In the interview, Dr. Sacha van Albada from the Jülich Institute of Neuroscience and Medicine (INM-6) talks about the results of her work, which were published in the journal Frontiers in Neuroscience.

Does it make any difference in the results whether you calculate a neural network classically on a supercomputer or with the neuromorphic hardware you have tested?

Generally, the computers yield slightly different results, simply because of the differences in the representation of numbers: so-called floating-point on classic supercomputers and fixed-point on SpiNNaker. This means that on today’s version of SpiNNaker, there are a fixed number of digits before and after the decimal point, and thus a limited accuracy with which the numbers can be represented. In addition, different random numbers are usually drawn on different systems, as it would be very inefficient to write out the large number of random numbers on one system and feed them into the other system. However, these differences also occur between simulations done on the same computer and are not important in the scientific sense. Only the statistical properties are important to us, and these are also well represented by the neuromorphic simulations.

What are the advantages over normal supercomputers?

SpiNNaker allows you to trade accuracy for less time and energy consumption. When network activity is relatively low and asynchronous, SpiNNaker delivers all spikes that transmit signals between neurons. But if the neurons on a computer core get many spikes at the same time, SpiNNaker has the possibility to drop spikes so that the simulation continues to run quickly ─ even if the dynamic equations are solved with slightly less precision.

The idea behind this is that the function of the brain is also robust against various types of ‘noise’ in neuronal activity and signal transmission, including the loss of spikes. In principle, communication between computer cores that is adapted to the brain also ensures that the communication between the nodes of the computer grows only linearly with the size of the network, although the number of possible communication paths increases quadratically. This makes SpiNNaker a promising candidate to quickly simulate much larger networks in the future than we have done so far.

In what way is the SpiNNaker system different from a conventional supercomputer regarding energy consumption and speed?

In order to compare energy consumption and speed between computers, you must first ensure that the results are equally accurate. However, it depends on the network and goal of the simulation how the accuracy of the results is measured. In our case it was important that the firing rates, synchronisation, and irregularity of the spikes were statistically equal.

Under this condition, the speed and energy consumption of both systems were approximately the same in the simulation phase in which the neural dynamics are calculated. Depending on the parallelisation of the simulations with the NEST simulator, with which we implemented the network on a conventional computer, NEST was slower or faster than SpiNNaker. The fact that both systems are now almost equally efficient means that the neuromorphic system can surpass the efficiency of the supercomputer or cluster computer with the next generation of SpiNNaker chips.

At the moment, we cannot yet simulate such large networks on SpiNNaker like we can on a supercomputer. We have achieved a breakthrough in our current study, however. Our network has reached the majority of connections per neuron. This means that a larger number of connections per neuron in larger networks is but limited, and thus they come into range of what can be simulated. As part of the European Human Brain Project, this is exactly what our next step will be: to port significantly larger networks to SpiNNaker.

When do you expect the first neuromorphic systems to come onto the market?

There needs to be a distinction made. The experimental systems like SpiNNaker, which are very close to biology, are not commercially available at the moment, rather the developers make them available to other research groups. On the other hand, there are more specialised neuromorphic systems for tasks such as image and speech recognition which are already on the market and already installed in some smartphones, for example.

Originalpublikation: van Albada Sacha J., Rowley Andrew G., Senk Johanna, Hopkins Michael, Schmidt Maximilian, Stokes Alan B., Lester David R., Diesmann Markus, Furber Steve B.
Performance Comparison of the Digital Neuromorphic Hardware SpiNNaker and the Neural Network Simulation Software NEST for a Full-Scale Cortical Microcircuit Model
Frontiers in Neuroscience (published online 23 May 2018), https://doi.org/10.3389/fnins.2018.00291

Tobias Schlößer