Resistors that work in a similar way to the body’s nerve cells can be used to build neural networks for machine learning.
Many machine learning models rely on increased processing power to achieve results, but this has high energy costs and generates large amounts of heat.
One proposed solution is an analog learning machine that acts like a brain, using electronics like neurons to become part of the model. However, these tools so far are not fast, small or efficient enough to provide the benefits of digital, machine learning.
At the Massachusetts Institute of Technology, Myrat Onen and his colleagues have developed nanoscale resistors that move protons from one terminal to another. It works a bit like a synapse, a connection between two neurons, where ions flow in one direction to transmit information. But these “artificial synapses” are 1,000 times smaller and 10,000 times faster than their biological counterparts.
Just as the human brain learns by reconfiguring connections between millions of interconnected neurons, machine learning models can run on networks of these nanoresistors.
“We’re doing things that are quite similar, like ion transport, but we’re doing it so fast now that it’s not possible in biology,” Onon said.
The resistor uses a strong electric field to transport protons at very high speeds without damaging or breaking the resistor itself, a problem experienced by previous solid proton resistors.
A practical analog learning machine would require a system containing millions of resistors. One admits it’s an engineering challenge, but the combination of all materials and silicon should make it easy to integrate with existing computing architectures.
“This is very high-speed, low-energy and efficient for what they have achieved technologically – it looks really impressive,” says Sergey Saveliev at Loughborough University in England. However, the fact that the device uses three terminals, instead of two, like human neurons, can cause some nervous systems to be difficult to control.
Pavel Borisov, also at Loughborough University, agrees that this is an effective technology, but points out that the protons come from hydrogen gas, which could prove difficult to store in the device when improving the technology.