Zurich – After ten years of hard work at IBM Research, in Zurich, researchers have created the first functional artificial neurons, which could pave the way for super-computers.
For the longest time, computer scientists have dreamed of neuromorphic (“brain-like”) systems: those that are inspired by, and function just like, a natural, biological brain. IBM Research has made a crucial breakthrough, which could enable the creation of super-dense computing systems that can emulate the human brain’s powerful parallel processing architecture.
The human-made neurons
The artificial neurons have a membrane made of phase-changing materials, which can change their states, from a crystalline, orderly form to an amorphous, disorganized one and back again. These different phases, also behave differently, the crystalline structure is a conductor, and the amorphous form is an insulator. In the case of the neurons’ membrane, the material used is the germanium-antimony-tellurium (GeSbTe or GST), which has been previously used as a key component in optical discs.
The material changes its form when exposed to heat. In the case of the neurons, it does so via electric spikes, which allows the artificial membrane to emulate the behavior of its biological counterpart, the lipid bilayer. In turn, this permits the neurons to imitate another bit of behavior from biological cells, its stochasticity.
The stochasticity is what permits every neuron to behave differently, allowing new ways to compute and interpret signals and data. The reason for this is that, with every electrical spike and subsequent phase change, the resulting phase is always slightly different – the stages change.
Why would they be useful
The artificial neurons are currently of 90 nanometers in size; the study comments on the possibility of reducing them to 14 nanometers. A nanometer in one billionth of a meter. Their diminutive size is what will allow creating compact systems of neurons.
The compact system will permit the artificial neurons to occupy less space than the alternatives we have now. IBM researchers have already done that by organizing hundreds of neurons into a ‘population’, and using them to process a broadband signal in a manner not too dissimilar from how biological neurons work.
The neurons created by IBM can sustain billions of switching cycles thanks to the materials they are made of. These neurons need less than five picojoule (a picojoule being a million millionth of a Joule) and less than 120 microwatts of energy to update, according to Computer World.
IBM has mentioned that they’re also working on complementary technology, namely synapses for the neurons, which is what permits biological neurons to pass an electric – or chemical – signal to another one, and a critical characteristic needed for the correct function of these artificial ones.
While it’ll be quite a few years before scientists create the first processor made of these artificial neurons, it is still a crucial development that moves humanity closer to a better future. Neuromorphic computer systems could process massive amounts of data ridiculously fast, and also be capable of detecting patterns and make data correlations in real-time.