IBM scientists create artificial neurons in claimed breakthrough for AI

Based on an alloy used in Blu-ray discs, the artifical neurons look promising as the basis for next-generation computing

IBM scientists together with colleagues at the Swiss university ETH Zurich have claimed a breakthrough in efforts to build a machine that computes like a human brain.

Using readily available and well-understood materials, the researchers have found a way to imitate the actions of neurons, which fire an electrical charge across a membrane where it is picked up by a synapse, with the potential for industrial production.

The human brain contains 100 billion neurons which are supported by many more glia cells. Each neuron may be connected to 10,000 other neurons creating as many as a quadrillion synaptic connections in total. The brain's memory capacity is of the order of terabytes and yet it only consumes about 20W of power and takes up a mere two litres of space.

This makes it a far more efficient processing engine than a classical computer, and because neurons function stochastically, meaning that there is a degree of randomness in their operation, the brain is very good at some learning and pattern matching-type operations that machines find difficult.

The artificial neurons developed by the scientists mimic the biological process of by which neurons transmit data to the synapses. Germanium antimony telluride, the basis of data storage on Blu-ray discs, changes phase when an electrical pulse is applied, becoming crystalline and less resistant and thus allowing the artificial neuron to fire. The team claims that this is the first time that an artificial 'integrate-and-fire' neuron has been created based on phase-change materials.

This alloy is relatively cheap, stable over billions of switching cycles and can be used to create artificial neurons at the kind of scales required for micro-circuits. What's more, the artificial neurons are capable of operating with a low power input, each one needing just five picojoules of energy to trigger it.

The IBM team managed to bring several hundred of these tiny artificial neurons together and used them to perform simple computational tasks.

Large populations of these nanoscale neurons could be used to create neuromorphic processors that are able to learn from their experience. Other use cases include IoT sensors and next-generation parallel computing.

"Populations of stochastic phase-change neurons, combined with other nanoscale computational elements such as artificial synapses, could be a key enabler for the creation of a new generation of extremely dense neuromorphic computing systems," said scientist Tomas Tuma.

The team is now working on scaling down their prototype devices and making them even more energy efficient.