Japanese researchers have developed a method to create conductive polymer wire connections (Developing polymer-inspired brains for artificial neural networks) between electrodes to make artificial neural networks that go beyond the limits of traditional computer hardware.
The development of neural networks to create artificial intelligence in computers was originally based on how biological systems work. These “neuromorphic” networks work on hardware that has no resemblance to the biological brain, which hinders performance.
Now, researchers from Osaka University and Hokkaido University plan to change that by creating neuromorphic “wetware.” Although neural network models have achieved significant success in applications such as image generation and cancer diagnosis, they still lag behind the general processing capabilities of the human brain. This is in part because they are embedded in software that uses traditional computer tools that are not designed for the millions of parameters and connections that are commonly required.
Neuromorphic wetware, based on memristive devices, can solve this problem. A memristive device is a device whose resistance is determined by the applied voltage and current history. In this method, electropolymerization is used to connect electrodes immersed in the first solution using conductive polymer wires.
The resistance of each lead is then adjusted using a small resistor, resulting in a memristive device. “The ability to build fast, energy-efficient networks using 1D or 2D systems has been demonstrated,” says lead author Megumi Akai-Kasaya. “Our goal is to extend this process to build 3D networks.”
The researchers were able to create polymer yarns from a polymer blend called “PEDOT:PSS,” which is highly conductive, reflective, flexible and stable. The 3D configuration of the upper and lower electrodes is first immersed in the first solution. PEDOT: PSS wires are formed between selected electrodes by applying a square voltage to these electrodes, mimicking the formation of synaptic connections through axonal guidance in the immature brain.
Once the wire is established, the characteristics of the wire, especially the activity, are controlled by using small voltage pulses applied to the electrode, which changes the electrical characteristics of the film surrounding the wires.
“The process continues and evolves,” says author Naruki Hagiwara, “and this feature is what allows the network to be trained, like software-based neural networks.”
The network created is used to demonstrate unsupervised Hebbian learning (that is, when synapses often fire together and strengthen their connections over time). In addition, the researchers can control the behavior of the wires so that the network can function. Spike-based learning, another method of neural networks that mimics the structure of neural networks, has also been demonstrated by controlling the diameter and conductivity of wires.
Then, by creating a chip with a large number of electrodes and using microfluidic channels to deliver the original solution to each electrode, the researchers hope to create a much more powerful system. In general, the method determined in this study is a big step to obtain neuromorphic wetware and to make the difference between the cognitive abilities of humans and computers.
Source: Osaka University
