As AI demands ever more power, researchers want to the brain for more efficient ways to process information. A brand new approach uses soft, flexible electronics to create artificial neurons that may mimic biological signaling and even directly interface with living neural tissue.
Researchers have long attempted to create so-called “neuromorphic” chips fabricated from artificial neurons that mimic the spiking behavior of their biological counterparts. But there are still wide gaps between how these devices and brains operate.
Real neurons within the brain display a wide range of activity patterns, which helps them encode and process information extremely efficiently. In contrast, most artificial neurons are carbon copies of one another with highly uniform spiking behavior, forcing neuromorphic chips to make use of tens of millions of those neurons to realize even modest functionality.
Now, a team from Northwestern University has designed a novel fabrication technique to create artificial neurons that mimic the complex signaling patterns present in the brain. The neurons’ output was so realistic that they successfully stimulated neurons in mouse brain tissue. More importantly, the approach could lay the groundwork for rather more energy efficient AI.
“Silicon achieves complexity by having billions of similar devices,” Mark Hersam, who co-led the research, said in a press release. “Every little thing is identical, rigid and glued once it’s fabricated. The brain is the other. It’s heterogeneous, dynamic and three-dimensional. To maneuver in that direction, we want recent materials and recent ways to construct electronics.”
The team created their artificial neurons, described in a paper in Nature Nanotechnology, by jet printing special electronic ink onto a versatile polymer. The ink accommodates nanoscale flakes of molybdenum disulfide, which acts as a semiconductor, and graphene, which serves as an electrical conductor.
The ink also accommodates a stabilizing polymer researchers typically burn off after printing to forestall it from interfering with the flow of current. However the researchers discovered that by leaving a few of it behind, they may introduce imperfections that lead to much more sophisticated signaling behavior.
Reasonably than completely burning the fabric away, they partially decomposed it. Then after they passed a current through the printed neurons, the polymer broke down further, but in an uneven pattern that created a conductive thread where current gets squeezed into a good channel.
This constricted pathway rapidly switches on and off, firing sharp voltage spikes that look loads just like the spikes present in real neurons. The device doesn’t just produce easy on-off pulses, but the whole lot from isolated spikes to sustained firing to rhythmic bursts, very similar to an actual neuron.
With just two of those printable neurons and a few basic circuit components, the researchers produced sophisticated spiking patterns. And crucially, they were in a position to tune the length and frequency of spikes to match the timing of biological motion potentials, which may very well be useful for applications like bioelectronic medicine or brain-computer interfaces.
To check whether or not they could transcend simply matching the numbers, the team worked with Northwestern neurobiology professor, Indira Raman, to hook up their artificial neurons to slices of mouse cerebellum and fire spikes into the tissue. The biological neurons fired in response, showing the synthetic signals were convincing enough to activate real neural circuits.
“You may see the living neurons reply to our artificial neuron,” said Hersam. “So, we have demonstrated signals that aren’t only the suitable timescale but in addition the suitable spike shape to interact directly with living neurons.”
While those capabilities may lead to some interesting applications, the researchers’ mainly hope the technology can reduce AI’s energy bill by mimicking the brain’s more efficient processing.
“To fulfill the energy demands of AI, tech corporations are constructing gigawatt data centers powered by dedicated nuclear power plants,” Hersam said. This will only scale up to now, when it comes to power and cooling, he said. “Nevertheless you have a look at it, we want to give you more energy-efficient hardware for AI.”
Given the long, tortuous path from lab bench to factory floor, it seems unlikely this technology can be making a dent within the industry’s power bill any time soon. Nevertheless it could lay the groundwork for a better approach to do computation in the longer term.