Atomically bio-plausible neuron toward complex neuromorphic applications

Neuromorphic computing, benefitting from its integration of computing with memory, enables highly efficient parallel-computing capabilities. While artificial intelligence chips are expensive due to their large area and power consumption, neuromorphic devices have shown energy efficiency and compatibility with complementary metal-oxide-semiconductor transistor technology.¹ Complex neuronal circuits with feedforward and feedback topologies are the foundation for nonlinear information integration and processing in the human brain. In addition, the nonlinear integration of neuronal signals as the basic functions of the human brain's nervous system is also essential to implement machine learning. However, artificial neurons still face the challenge of nonlinearly integrating feedforward and feedback signals. It is crucial to develop bio-plausible neurons capable of those functions, including nonlinearity and integration of excitatory and inhibitory postsynaptic signals. Writing in Nature Nanotechnology, G. S. Syed and coworkers recently reported a major step toward bio-plausible optomemristive feedback neurons, enabling the simultaneous existence of separate feedforward and feedback paths within a neural network.²

The authors designed a delicate capacitor-like device with a 2D vertical heterostructure in which WS₂/MoS₂ and graphene served as the neuronal membrane and soma (Figure 1B), respectively. Generally, trapped electrons and holes in the WS₂/MoS₂ heterostructure recombine upon a positive back gate voltage (Figure 1A). The conductance state of p-doped graphene would further increase, representing an excitatory operation. In this work, the electron-hole carriers in the WS₂/MoS₂ heterostructure are easily separated upon illumination (Figure 1C), and the electrons are injected into graphene. The Fermi-level movement toward the Dirac point decreases the conductance of graphene, having an inhibitory effect. Specifically, graphene's gradual conductance change can be separately modulated through electrical and optical means (Figure 1D), mimicking excitatory and inhibitory functionalities. 2D memristors have been investigated to emulate leaky-integrate-and-fire feedforward neurons.³ The synergistic effect of both input signals mimics a competitive neuron and enables the simultaneous existence of separate feedforward and feedback paths within the neural network.

The winner-take-all (WTA) neural network is a critical computational model for artificial neural networks, which can be used to implement unsupervised competitive learning and cooperative learning. The traditional memristors make it difficult to separately process feedforward and feedback neuronal signals, necessitating peripheral circuits or software to mimic inhibition behavior. The developed optomemristive feedback neuron can respond to both electrical and optical stimulation and broadcast inhibitory signals on neighboring neurons and nonlinear integrated neuronal signals. Therefore, the authors further created a WTA neural network to demonstrate its superiority (Figure 1E) in which WTA neurons comprise the output layer of the neural network, acting as a rectifier activation function. The WTA neural networks implement the neuronal signal accumulation and activation tasks and demonstrate the potential for unsupervised competitive learning and cooperative learning.

It has been demonstrated that 2D materials facilitate the construction of multi-terminal memtransistors for complex neuromorphic functions.⁴ Neuromorphic perception devices such as the artificial retina could be emulated by leveraging their excellent photoresponsivity.⁵ This work greatly advances the hardware implementation of bio-plausible neuromorphic devices and highlights important routes to solve complex tasks by developing WTA neuronal networks. The charge trapping effect is essential to realizing the negative photoresponsivity of 2D materials and the ability to process feedback neuronal signals. Treating the SiO₂/Si substrate or directly depositing the Al₂O₃ layer are two common approaches to introduce the trapping effect, which is an efficient way to realize negative photoresponsivity toward optomemristive feedback neurons. The unique device structure and vdW heterostructure innovation are also responsible for those fascinating properties.

The spatiotemporal complexity of the human brain cortex and its neural networks is the basis for the higher intelligence of humans. Despite the progress made in this work, there is a vast gap between neuromorphic computing and the human brain due to their differences, including structure, working mechanisms, and scale. In addition, we consider that a brain-like design in terms of working mechanisms, device connectivity complexity, and scale is an effective and even necessary way to achieve complex neuromorphic applications. Memristive devices with neuron-like structures and memristive mechanisms are greatly needed to faithfully emulate biological neuron functions.

Song Hao: Conceptualization, Formal analysis, Funding acquisition, Project administration, Validation, Writing—review and editing; Yanfang Niu: Data curation, Funding acquisition, Resources, Writing—review and editing; Shancheng Han: Data curation, Writing—original draft.

The authors declare no conflicts of interest.

Ethics approval was not needed for this study.