Introduction
Neuromorphic computing holds tremendous potential, yet the hardware costs remain prohibitively high. So, how can we overcome this challenge? The latest research from the laboratory combines memristor pulse generators with electrochemical transistor state control units to efficiently realize reconfigurable bio-inspired neurons, capable of generating complex spiking behaviors. This breakthrough pushes forward the integration of artificial intelligence and brain science!
On February 5, 2025, the Pan Gang and Lin Peng research teams from Zhejiang University's State Key Lab of Brain-Machine Intelligence, the College of Computer Science, and the Liangzhu Laboratory, in collaboration with Researcher Kong Wei from Westlake University, published a paper titled "Bio-plausible reconfigurable spiking neuron for neuromorphic computing" online in Science Advances. The study leverages the efficient pulse generation capability of MOTT memristors and the precise state control of ECRAM to develop a new type of bio-inspired neuron. This neuron mimics the functioning of biological neurons, offering high-efficiency and flexible signal processing while using fewer hardware resources and at a lower cost. This achievement addresses the complexities and high costs associated with traditional bio-inspired neuron designs and opens up new perspectives for building more intelligent neuromorphic computing systems. It also offers fresh directions for advancing the intersection of artificial intelligence and brain science.
Neuromorphic Computing aims to simulate the dynamic behavior of neurons and synapses in the brain to enhance the computational efficiency and cognitive abilities of artificial intelligence systems. It is a hot topic in current research. Artificial neurons, as the fundamental building blocks of neuromorphic computing systems, draw inspiration from the diverse firing patterns of biological neurons and play a key role in improving the information processing capabilities of brain-like systems.Existing bio-inspired neuron models based on traditional CMOS technology, such as the Hodgkin-Huxley (H-H) model, feature multiple spiking modes that can be applied to various information processing scenarios, such as biological rhythmic activities and sensory information encoding. However, because CMOS devices lack intrinsic neural dynamics, constructing circuits to build these neurons is complex, making it difficult to implement and apply directly in hardware systems.On the other hand, simplified neuron models, such as the Leaky-Integrate-Fire (LIF) model, though widely used in the design of neuromorphic computing systems, have a single spiking mode and cannot fully replicate the rich spiking behaviors of biological neurons. This limitation restricts the cognitive performance of neuromorphic computing systems.
Figure 1: Heterogeneous integrated neuron circuits based on MOTT memristors andECRAM
Figure 2: Reconfigurable neurons can switch between different spiking modes through ECRAM resistance control
This research, inspired by classic fast-slow neuron circuit analysis methods, introduces a new type of bio-inspired brain-like neuron circuit based on MOTT memristors and ECRAM for heterogeneous integration (Figure 1). By utilizing the efficient pulse generation capability of MOTT memristors and the precise state control of ECRAM, this design achieves a bio-mimicry that is highly functional, adjustable, and cost-effective in terms of circuit resources. Unlike previous studies that altered neuron spiking patterns by changing input stimuli or replacing circuit components, this work is the first to achieve diverse neuron firing pattern modulation simply by programming the device states (Figure 2). This highlights the advantages of brain-like devices, such as low hardware overhead and high-density integration, in neuromorphic computing systems, providing new hardware support for more efficient artificial intelligence systems.
More biologically plausible neuron firing patterns play a crucial role in neural network information processing. For example, cluster-firing neuron models in spiking neural networks allow for more precise control of spike timing, achieving better classification accuracy with smaller time steps, thus optimizing both accuracy and operational efficiency. Future research will delve deeper into leveraging new device heterogeneous integration technologies to develop more efficient and intelligent neuromorphic computing chips and systems, with applications in emerging fields such as brain-computer interfaces and brain-inspired computing, facilitating the integration and mutual advancement of artificial intelligence and brain science.
Doctoral students Xiao Yu and Zhang Bihua from Zhejiang University’s College of Computer Science and Technology, and undergraduate Liu Yize from the College of Information Science and Electronic Engineering, are the first authors of this paper. Liu Yize joined the project through Zhejiang University's "Pioneering Talent Development Program," a platform that selects excellent mentors from national key laboratories and allows for mutual selection with well-rounded undergraduates to engage in research projects or other forms of scientific training, providing undergraduates with the opportunity to participate in high-level research projects. Professor Pan Gang, Researcher Lin Peng from Zhejiang University’s State Key Lab of Brain-Machine Intelligence , and Researcher Kong Wei from Westlake University are the corresponding authors. Additionally, Dr. Chen Peng (postdoctoral fellow), Researcher He Enhui, PhD student Huo Wenju, Associate Professor Ma De, Researcher Zheng Qian, Professor Tang Huajin from Zhejiang University, and Associate Researcher Zhang Xumeng, Researcher Jiang Hao from Fudan University also made significant contributions to this study. This research was supported by the Key R&D Program Major Project "Neuron Computer Systems with Integrated Perception and Computing," the National Outstanding Youth Science Fund, the Zhejiang Provincial Natural Science Foundation Major Project, and the Zhejiang University Zijin Program.
Original paper link：
https://www.science.org/doi/10.1126/sciadv.adr6733
