A team from the University of Southern California (USC) has developed a diffusive memristor that can simulate the electrochemical behavior of human brain neurons and produced a prototype chip that integrates multiple neurons. Relevant results were p...
A team from the University of Southern California (USC) has developed a diffusive memristor that can simulate the electrochemical behavior of human brain neurons and produced a prototype chip that integrates multiple neurons. Relevant results were published in "Nature Electronics", demonstrating an ion-driven brain-like structure that can realize learning and memory in hardware, opening up a new path of high energy efficiency for future AI chips.
The team is led by Joshua Yang, a professor of electrical and computer engineering at USC, who is famous for his research on artificial synapses more than a decade ago. This achievement allows electronic components to physically reproduce the firing behavior of neurons, laying the foundation for the hardware implementation of AGI.
First of all, let’s understand what is a “diffusive memristor”? It is a special electronic component that can simulate the firing and recovery behavior of neurons. Its conductive state changes dynamically as ions "diffuse" and "reflow" in the material.
In this experiment, Yang's team used silicon oxide (SiOx) and silver nanoparticles (Ag NPs) to make a flexible memristor. When stimulated by external force or voltage, silver ions will diffuse in the oxide layer and form a conductive channel, generating a "current spike" like a neuron.
This device can discharge itself when subjected to mechanical pressure and change the discharge pattern according to the intensity or frequency of the pressure, simulating the chemical dynamics of "synaptic transmission" in the brain. Research shows that it only needs the size of a transistor to reproduce the behavior of human brain neurons. It is a physical precision that is difficult to achieve by existing simulation systems. It also means that the number of transistors required on the chip will be greatly reduced, making the overall structure more compact and efficient.
Yang pointed out that although the traditional electronic chip-based learning mechanism is fast, it consumes high energy and lacks stability. In comparison, ion-based hardware learning is closer to the way the human brain operates and can achieve stable responses at lower energy consumption.
However, the research team also admitted that there are still practical challenges. Since the silver ions (Ag⁺) used in the experiment are easy to diffuse in the wafer and are not completely compatible with the current CMOS process, it is difficult to directly introduce it into commercial wafer production. In addition, although this research successfully reproduced the firing response of neurons, it is still in its early stages and has not yet entered the development of large-scale systems.
In the future, the team will continue to explore other alternative ionic materials and structural designs, hoping to gradually advance the practical process of "brain-like chips" while retaining energy efficiency.
Artificial neurons developed by USC team replicate biological function for improved computer chips Further reading: Shiitake mushrooms transformed into memory resistors and no longer want to be used as soup ingredients
