Paul Marks writes in New Scientist that the material that lets us record on DVDs has a far more tantalising property: it can mimic the nerve cells of the brain and the junctions between them. The discovery could lead to the development of brain-like computers that, crucially, operate at ultra-low power levels.
A brain-like computer is one that can learn and adapt without external programming. Such an ability would allow machines to become far better at tasks like face and speech recognition. They could also process and store data in the same location – just as nerve cells do. Conventional computing loses efficiency by keeping these functions separate.
Now two research groups have built artificial nerve cells, or neurons, and synapses – the junctions between them – using an alloy known as GST, an acronym of the symbols for its components: germanium, antimony and tellurium.
In the UK, David Wright and colleagues at the University of Exeter have created a GST neuron (Advanced Materials, DOI: 10.1002/adma.201101060), while at Stanford University in California, Philip Wong’s group have created a nanoscale electronic synapse. The junction even mimics the way synapses can change their connection strength (Nano Letters, DOI: 10.1021/nl201040y).
GST is known as a “phase-change” alloy, because of its ability to change its molecular structure from a crystalline to a disordered amorphous “phase” when heated. In DVDs, this allows binary 0s and 1s to be recorded and then read by a laser.
But GST can do more than store two states. Different areas within a tiny spot of GST can be crystalline or amorphous to differing degrees, which means it can store information across a much wider range of values than simply 0 or 1. This is important because it is a build-up of input signals that makes a real neuron “fire” when it reaches a certain threshold.
Wright’s neuron is able to mimic this threshold firing because GST’s electrical resistance drops suddenly when it moves from its amorphous phase to the crystalline. So incoming signals in the form of pulses of current are applied to the artificial neuron – and it is deemed to have fired when its resistance plummets.
GST’s talents don’t end there. When a real neuron fires, the signal’s importance to the next neuron it arrives at is set by the strength of the synapse connecting them. In nature, this strength is adjusted in a process called spike-timing-dependent plasticity (STDP): if the first neuron repeatedly fires before the second, the synapse’s strength increases, but if the second fires first, its strength decreases.
Duygu Kuzum, a member of the Stanford team, says GST’s ability to change its resistance has allowed them to program it to dynamically modify the strength of the nanoscale artificial synapses they have built – just like STDP. This lets them prioritise which neural signals are most important to any given task.
At just 75 nanometres across, the artificial synapse may offer the low power sought for brain-like computers, says Kuzum. The team’s calculations suggest a system with 1010 synapses would consume just 10 watts – compared with the 1.4 megawatts needed by a supercomputer to simulate just 5 seconds of brain activity.
“Phase-change devices may indeed capture the right essence of the behaviour of the brain,” says Steve Furber of the University of Manchester, UK, who is building a brain-like computer from conventional microprocessors. “But it has a very long way to go. I’ll be interested when they can make 100 million of them on a chip for next to nothing.”