Researchers have been trying to build artificial synapses for years in hopes of approaching the unrivaled computational performance of the human brain. A new approach has now succeeded in designing them 1,000 times smaller and 10,000 times faster than their biological counterparts.
Despite the overwhelming success of deep learning over the past decade, this brain-inspired AI approach faces the challenge of being run on hardware that bears little resemblance to real brains. This is an important part of why a human brain weighing only three pounds can do new activities in seconds using the same amount of energy as a light bulb, while training larger neural networks takes weeks, megawatt-hours of electricity, and racks of specialized transformers.
This is sparking growing interest in efforts to redesign the underlying hardware that the AI runs on. The idea is that by building computer chips whose components act more like natural neurons and synapses, we may be able to get closer to the extreme space and energy efficiency of the human brain. The hope is that these so-called “neuromorphic” processors could be much better suited to running AI than today’s computer chips.
Now MIT researchers have shown that an unusual artificial synapse design that mimics the brain’s reliance on moving ions could actually significantly outweigh biological ones. The key breakthrough was the search for a material that could tolerate extreme electric fields, which greatly improved the speed at which ions could move.
“The speed was certainly surprising “, Murat Onen, who led the research, he said in a press release. “Normally, we wouldn’t apply such extreme fields to all devices, so as not to turn them to ash. But instead, protons [which are equivalent to hydrogen ions] it ended up moving at tremendous speeds through the stack of devices, specifically a million times faster than what we had before. “
While there they are a variety of approaches to neuromorphic engineering, one of the most promising is analog computing. This seeks to design components that can leverage their internal physics to process information, which is much more efficient and straightforward than performing complex logic operations as conventional chips do.
So far, much research has focused on the design of “memristors”, electronic components that control the flow of current based on the amount of charge previously flowed.and through the device. This mimics how connections between biological neurons increase or decrease in strength depending on how often they communicate, meaning that these devices could in principle be used to create networks with similar properties to biological neural networks.
Perhaps unsurprisingly, these devices are often built using memory technologies. But in a new one card inside Science, MIT researchers argue that components optimized for long-term information storage are actually unsuitable for performing the normal state transitions needed to continuously adjust connection strengths in an artificial neural network. This is because the physical properties that ensure long retention times are generally not complementary to those that enable high-speed switching.
Instead, the researchers designed a component whose conductivity is regulated by the insertion or removal of protons into a phosphosilicate glass (PSG) channel. To some extent, this mimics the behavior of biological synapses, which use ions to transmit signals across the gap between two neurons.
However, this is where the similarity comesyes end. The device has two terminals which are essentially the input and output of the synapse. A third terminal is used to apply an electric field, which stimulates protons to move from a reservoir into the PSG channel or vice versa depending on the direction of the electric field. More protons in the channel increase its resistance.
Researchers It came up with this general design in 2020, but their previous device used materials that were not compatible with the chip design processes. Most importantly, the switch to PSG has greatly increased the switching speed of their device. This is because the nano-sized pores in its structure allow protons to move very quickly through the material and also because it can withstand very strong electric field pulses without degrading.
More powerful electric fields give protons a huge speed boost and are critical to the device’s ability to bypass biological synapses. In the brain, electric fields must be kept relatively weak because anything above 1.23 volts (V) causes the water to produceS. most cells to divide into hydrogen and gaseous oxygen. This is largely why neurological processes occur on the millisecond scale.
In contrast, the MIT team’s device is capable of operating up to 10 volts with pulses of just 5 nanoseconds. This allows the artificial synapse to operate 10,000 times faster than its biological counterpartS.. What’s more, the devices are only nanometers wide, making them 1,000 times smaller than biological synapses.
Experts said New scientist that the device’s three-terminal configuration, as opposed to the two found in most neuron models, could make it difficult for certain types of neural networks to run. The fact that protons must be introduced using hydrogen gas also presents challenges when technology is stepped up.
There is a long way to go from a single artificial synapse to large networks capable of performing serious information processing. But the exceptional speed and small size of the components suggest that this is a promising direction in the search for new hardware that can match or even exceed the power of the human brain.
Image credit: Ella Maru Studio / Murat Onen