Yet another tech step-up achieved by IBM in creating the world’s first simulated phase change neurons. Their research team in Zurich has invented first artificial nanoscale stochastic phase change neurons. Already approx. 500 no. of these neurons have been produced which processes signal just like the natural neuromorphic way.
To them who have no idea what a neuron does, it is a specialized cell which transmits nerve impulses. One of the key effects of these kinds of neurons is-they are capable of firing at a very high speed with a low budget energy requirement. Just like biological neurons these neurons also produces stochasticity-their ability to produce slightly different, random result.
IBM’s artificial neuron has inputs (dendrites), a neuronal membrane (soma, nucleus), and an output (axon). There is also a feedback-like system called back propagation, which ensures the strength of these spikes remains maintained.
The main difference is in the neuronal membrane. In a real neuron there would be a lipid bilayer coating, which acts as a resistor and capacitor. It resists conductance, but with time gains enough electricity to produce its own spike of electricity-which is then carried to axons to other neurons-and so on and on.
In IBM’s artificial neuron, the membrane is substituted by a small square of germanium-antimony-tellurium (GeSbTe or GST). GST, which is said to be the active ingredient in rewritable optical discs, is a phase change material, which gives it a freedom to live in any form i.e. amorphous or crystalline and can easily switch between them. Here in GST’s case the amorphous phase is electrical insulator, while the crystalline phase conducts.
During the start of the GST cycle, it remains in amorphous phase. Then as spikes hits it from inputs, it gains electricity and slowly transforms to crystalline form which allows it to conduct electricity. After an arbitrary refractory period (a resting period where something doesn’t responds to stimuli) the GST bounces back to its amorphous stage and the process begins again.
The uses of these materials seem to be of wide range as it lasts long (trillions of switching cycles). Also they can be fabricated/integrated on leading-edge nodes. The phase change devices used in this experiment is measured to be very small and efficient-squares that are about 100nm in dimension.
Secondly, it can also be used in computer technology where sensory information is required and the computer would mostly behave as biological neuron while decision-making. We can hope to see more massive, efficient computer designs in future with the help of this research.
Designing and fabricating thousands of these neurons into single chip is what the next target is. Can’t stop here though, we can think of developing some software that would actually work on these chip’s neuromorphosity.