What makes IBM’s achievement notable from other researches is that their artificial neurons are built out of famous materials that can scale down to a few nanometers but can still activate with low energy, points out Ars Technica. Also important is the neurons’ stochasticity, which means their ability to always generate slightly different, random results, like biological neurons. Organic neurons have membranes acting as signal gates that take a definite amount of energy to absorb. In IBM’s neuron, the membrane is replaced with a small square of Germanium-Antimony-Tellerium (GeSbTe or GST), a common component in optical discs. GST is a phase-change material, which also happens to be the main active ingredient in rewritable optical discs. Once the GST is heated enough, it changes its physical phase from an amorphous insulator to a crystalline conductor. In the case of GST, its amorphous phase is an electrical insulator, while the crystalline phase conducts. This means that when the artificial membrane is hit with sufficient electricity to change into its crystal phase, signal passes through and then it resets to its amorphous one. However, the scientists required artificial neuron to have another trait of its organic counterpart: stochasticity, or some randomness in when signals will fire. IBM says its neurons are able to attain this because its GST membranes are always slightly different after each reset, which in turn causes the crystallisation process to be different. As a result, the engineers are never sure exactly when each artificial neuron will fire, which in turn allows them to achieve things suddenly that they could not if their results were perfectly foreseeable. Computers that imitate the effective and similar kind of processing design of organic brains can be developed by scientists with the help of these neurons and its style of approach to processing sensory information and decision making could be applied, suggests Ars Technica. However, as Ars Technica points out that developing these neurons is the simpler part but writing software that actually makes use of the chip’s neuromorphosity will altogether be another challenge. Source: Ars Technica
