The future of neural network computing may be a little murkier than we expected.
A team of physicists has successfully developed a processor “ionic circuit” that uses the movement of charged atoms and molecules in aqueous solutions rather than electrons in solid-state semiconductors.
The device could be the next step in brain-like computing, they say, because this form is closer to how the brain communicates information.
In a new paper, a team led by physicist Woo-Bin Jung at the John A. Paulson School of Engineering and Applied Sciences (SEAS) at Harvard University says, “Ionic circuits in aqueous solutions utilize ions as charge carriers for signal processing. I am aiming for,” he said.
“Here we report an aqueous ionic circuit… Demonstration of a functional ionic circuit capable of analog computation is a step towards more advanced aqueous ionics.”
Critical to signaling in the brain is the movement of charged molecules called ions through liquid media. Replicating the brain’s incredible processing power is extremely difficult, but scientists wondered if a similar system of pushing ions into an aqueous solution could be used for computing.
In this case, it will be slower than traditional silicon-based computing, but it could have interesting advantages.
For example, ions can be made from different molecules, each with different properties and can be used in different ways.
But before that, scientists have to prove it works.
This is what Jung and others have been working on. First, we designed the function of the ion transistor, a device that switches signals and supplies power. Recently, he succeeded in combining hundreds of such transistors to function as an ion circuit.
A transistor consists of a small disc-shaped electrode in the center, surrounded by two concentric ring-shaped electrodes, called a “bull’s eye” electrode arrangement. It interacts with aqueous solutions of quinone molecules.
Applying a voltage to the central disk generates a current of hydrogen ions in the quinone solution. Meanwhile, two ring electrodes adjust the pH of the gating solution to increase or decrease the ionic current.
This transistor performs a physical multiplication of the disk voltage with a “weight” parameter set at the gate of the ring pair to produce a response such as ion current.
However, neural networks rely heavily on a mathematical operation called “matrix multiplication” to perform multiple multiplications.
Therefore, the research team arranged 16 × 16 transistors and designed them so that they can perform arithmetic multiplication, and created an ion circuit that can perform matrix multiplication.
“In artificial intelligence neural networks, matrix multiplication is the most common computation,” says Jung. “Our ionic circuit performs matrix growth in water in an analogue manner that relies entirely on electrochemical mechanics.”
Of course, this technique has important limitations. Since the 16 currents cannot be solved separately, the operations had to be done sequentially rather than simultaneously, slowing down the already slow technology significantly.
But its success is a step towards more advanced ion computing. Only by looking at the problem can the solution be found.
The next step is to introduce a wider range of molecules into the system and see if the circuit can handle more complex information.
“To date, we have only used three or four ionic species, such as hydrogen ions and quinone ions, to enable gating and ion transport in aqueous ion transistors,” says Jung.
“It will be very interesting to see how we can take more diverse ion species and use them to enrich the information content to be processed.”
The ultimate goal, the researchers say, is not to compete with or replace electronics, but to complement them, perhaps in the form of a hybrid technology that does both.
Leave a Reply