As artificial intelligence hardware advances toward higher efficiency and greater intelligence, enabling individual physical devices to perform more complex computations has become a key challenge. Recently, a research team from the School of Integrated Circuits and the Institute of Artificial Intelligence at Peking University published research in National Science Review, reporting a "locally active memristive oscillator" based on vanadium oxide (VO₂). When biased at the edge of chaos, the device exhibits rich dynamic behaviors controlled by tiny injected signals and possesses exceptional frequency-domain information processing capabilities, offering a breakthrough for dynamic neuromorphic computing.
The paradigm shift from "simple switches" to "dynamic computing"
Most memristors are often used for static and linear computing tasks, such as vector-matrix multiplication, with nonvolatile and locally passive dynamics. In contrast, memristors with "locally active" characteristics can amplify small fluctuations, generate self-oscillations, and produce complex responses, making them more similar to biological complex neurons. The team developed a well-fabricated VO₂ device and constructed a concise yet accurate thermodynamic compact model based on the theory of "local activity". It can unify material physical nonlinearity, circuit dynamics, and information processing capabilities for the first time. This is like neurons in the brain that can generate different responses based on the frequency of input signals without the need for complex circuit structures
Edge Chaos: The Golden Zone of Computing Power
The 'edge of chaos' region means locally stable and locally active. It means a special balance region which maintains stability while amplifying fluctuations. When the device operates in the edge chaotic region, its complex behavior can be controlled. In experiments, the team biased the VO₂ device at the "critical edge of chaos" state and injected small signals of different frequencies. The device exhibited various dynamic modes such as frequency division, random spiking, and frequency locking—all accurately predicted and simulated by the theoretical model. Furthermore, the team directly fed speech signals into the oscillator and utilized its frequency-sensitive properties to successfully classify spoken digits. Surprisingly, the dynamic response of just a single device achieved a classification accuracy of 92%, comparable to a two-layer convolutional neural network, with extremely low energy consumption.
Single device, neural network level performance
They demonstrated that a single VO2 device operating in the edge of chaos region can perform frequency domain feature extraction and achieve speech recognition, with performance equivalent to a two-layer convolutional neural network and reducing about 104 multiplication operations. More importantly, this process does not require additional computing resources, as frequency feature extraction is a natural result of the physical dynamics of the device itself.
"This work is not only an advance in device physics but also demonstrates a new computing paradigm," said corresponding authors Professor Yuchao Yang. "We have systematically verified, for the first time in experiments, the information processing advantages of locally active memristors at the edge of chaos. They can directly extract frequency features of signals without complex preprocessing, greatly simplifying system architecture and reducing power consumption."
This research provides an important theoretical and experimental foundation for building next-generation, energy-efficient neuromorphic chips with dynamic learning capabilities. It marks a paradigm shift from traditional "locally passive" devices to "locally active" devices and is expected to play a significant role in speech recognition, real-time signal processing, edge computing, and other fields.
Future Prospects
The research team in Peking University plans to apply this technology to a wider range of scenarios, including real-time speech recognition, sensor data processing, and more efficient neuromorphic computing systems. They believe that this physical dynamic computing method will provide new ideas for the next generation of computing devices, enabling computers to process information efficiently and flexibly like biological brains.
