Artificial intelligence and high-performance computing are driving up the demand for massive sources of energy. But neuromorphic computing, which aims to mimic the structure and function of the human brain, could present a new paradigm for energy-efficient computing.
To this end, researchers at Lawrence Livermore National Laboratory (LLNL) created a droplet-based platform that uses ions to perform simple neuromorphic computations. Using its ability to retain short-term memory, the team trained the droplet system to recognize handwritten digits and play tic-tac-toe. The work was published in Science Advances.
The authors were inspired by the human brain, which computes with ions instead of electrons. Ions move through fluids, and moving them may require less energy than moving electrons in solid-state devices.
"Think about what you had for breakfast," said LLNL scientist and senior author Aleksandr Noy. "It's not a lot of energy, but you are able to do pretty sophisticated computing and information processing tasks."
Using that idea as a base, Noy and his colleagues built a soft, flexible and small platform. Their device is relatively simple: two droplets of salt water, their periphery lined with lipids - fat-like compounds that don't dissolve in water. The two droplets are suspended in oil, where they touch and form a bilayer (two layers of lipids) that mimics a cell membrane. An electrode is inserted into each droplet and used to apply a voltage. The team then measures the current response of the droplet pair.
That response - the current flowing through the droplets - is the key to computing. In a system with no memory, the response to a specific spike of input voltage would be the exact same current, every time. But in the droplet system, the researchers observed behavior associated with memory: the device produced a slightly different current depending on the voltage that was applied previously.
The team illustrated this behavior with an example reminiscent of Pavlov's dog experiment. Normally, high voltages result in high current outputs. By giving the droplet system repeated training spikes of low and high voltages, they were able to observe high current outputs at low voltages. In other words, they trained the "dog" (AKA the droplets) to salivate (output high current) not only when presented with food (high-voltage input), but also when a bell is rung (low-voltage input).
"It's pretty fascinating how such a simple object can perform these functions," said Zhongwu Li, an LLNL postdoc and the first author of the paper.
With a solid understanding of the droplet's memory capabilities, the authors were able to collaborate with computer science researchers at the University of Southern California and Google Research to use reservoir computing algorithms to teach it to recognize handwritten digits and play tic-tac-toe.
For the handwritten digits, they fed the droplets a voltage "code" for each pixel in the image. Because the droplets have memory, each code led to a different output current. During an initial training stage, the team mapped that output to the correct numeral. With that mapping in place, the droplet was able to identify further handwritten digits.
The droplets then faced off against an ideal, standard computer in games of tic-tac-toe. The moves were again input as voltage codes into the droplet, and the output was mapped to what move the droplet should make next. After training, the droplet system was able to tie the standard computer consistently.
This proof-of-principle demonstration is not competitive with today's computer chips, which are much faster and much more sophisticated, but the authors emphasized that it needs to be explored for future energy-efficient computing technology.
"I don't think any of us will be replaced by droplets any time soon," said Noy. "But it is cool that you can teach a droplet to play a board game."
