Novel brain‐inspired optomemristive feedback neuron for neuromorphic computing

Traditional computing architectures based on complementary metal-oxide semiconductor technology suffer from von Neumann computing bottleneck,1 resulting in poor computing efficiency and a huge energy consumption. To surpass the limits of conventional computation, scientists have begun to imitate the computational behavior of the human brain.2 With the advantages of highly parallel computing, high error tolerance and low power consumption, the human brain and its neural systems have inspired the rapid development of novel neuromorphic computing hardware.3 There are ∼86 billion neurons in the biological neural system. Neurons can govern the membrane potential for associative learning, memory, and information processing, with important roles in brain-inspired neuromorphic computing. Therefore, constructing artificial neuron via electronic devices is key to the realization of neuronal dynamics in the human brain. Different types of memristive neurons have been reported recently, such as phase-change memory, Mott insulators, magnetic memory, diffusive memristors and ferroelectric memory. The integrate-and-fire neuron function and spiking neural networks could be simulated based on the integration characteristic of these artificial neurons. Besides the characteristic of integration, nonlinearity is another necessary characteristic in neuronal emulation, especially for integrating the datastream during neuromorphic computing. However, the realization of nonlinear integration of excitatory and inhibitory postsynaptic potentials has not been reported in above artificial neurons. It is in urgent need to develop a novel artificial neuron with both nonlinear and integrated capabilities for high-efficiency computing. The research team of Harish Bhaskaran proposed an atomically thin optomemristive feedback neuron using a stack of MoS2, WS2, and graphene (Figure 1).4 The heterojunction of MoS2/WS2 acts as a neural membrane, and the graphene acts as neural soma. Different from traditional artificial neurons, the proposed two-dimensional (2D) neuron device could exhibit a rectified-type of nonlinearity in its output characteristics without the need for additional circuitry and software. The 2D optomemristive neuron shows great potential in winner-take-all learning (WTA) computational tasks and unsupervised learning, which provide guidance for atomic-scale rectified and nonlinear optoelectronic neurons. Schematic of a two-dimensional optomemristive feedback neuron for winner-take-all learning computations. The key performance of device is based on the combination and broadcast of electrical excitatory signals and optical inhibitory signals, which could be used for nonlinear and rectified integration of information in neuromorphic computing. Under light illumination, electron-hole pairs could be induced and separated by the intrinsic field in transition metal dichalcogenides. The electrons transit from the heterojunction of MoS2/WS2 to graphene, and the holes transmit to the silicon substrate, resulting in a decrease in channel conductance and inhibitory effect. Under positive electrical stimuli, the carrier concentration decreases due to the recombination of trapped holes and electrons in graphene, resulting in an excitatory effect. Thus, when an external electric field is applied at the columns and light stimuli in free space, the 2D optomemristive neuron can exhibit excitatory and inhibitory behaviors, respectively. Because of the unique photoelectric modulation characteristics of 2D optomemristive neurons, unsupervised competitive learning for autonomous intelligence has been successfully implemented. The constructed two-layer competitive neural network consists of input-layer neurons and optomemristive neurons, in which input-layer neurons provide feed-forward signals to the optomemristive neurons for neuronal firing. The WTA computational model is the basis of competitive learning, where neurons compete with each other for activation. The inhibitory effect of WTA neurons allows the neural network to finish learning without back-propagation. During the process of unsupervised learning, the winner neuron activates itself for the highest activation, and the synaptic weights finish the update. The clustering flowers task is then successfully solved by the WTA neurons. Furthermore, this study demonstrates cooperative learning using the WTA mechanism and global optical modulation of optomemristive neurons. The core part of cooperative learning is that the weight update during learning iteration occurs not only in the winner neuron but also in selected neighboring neurons. The synaptic weights of WTA neurons evolve from random states to optimal states by cooperating with other output neurons. Moreover, the optical signal is spatially constrained by following the rule of P(r) ∝ P0exp (−Ωvr3), which ensures that the optomemristive neurons within a certain area update the weights via an optical signal threshold. Based on the reconfigurable softness and in-memory updates of optomemristive neurons, the combinatorial optimization of the traveling salesperson problem was solved by the cooperative learning optomemristive neural network. In summary, optomemristive neurons exhibit a rectified-type of nonlinearity and integration capability of excitatory and inhibitory potentials for WTA computing, including competitive learning and cooperative learning. Different from the conventional charge-based photodetector and thermal-heater-based memristors,5 optomemristive neurons show great potential for high-efficiency neuromorphic computing. Owing to the design of an atomically thin heterojunction (MoS2/WS2/graphene), optomemristive neurons could be activated by global optical inhibitory and local electrical excitatory signals. These key properties of 2D optomemristive neurons ensure that the WTA computation can be implemented by atomically thin electronic devices. From our perspective, traditional integrated circuits face the problems of size reduction and von Neumann computing bottleneck, which inspire the requirement for novel 2D optomemristive neurons with an atomic thickness and high-efficiency computing capability. The 2D optomemristive neurons proposed by Harish Bhaskaran could considerably promote the development of high-density and small-sized neuromorphic computing systems. Although detailed parameters of optomemristive neurons require further research for system-level application, we believe the proposed atomically thin optomemristive feedback neurons provide a new approach to hardware machine learning. Jialin Meng: Conceptualization; visualization; writing—original draft. Lin Chen: Review and editing. Tianyu Wang: Review and editing. David Wei Zhang: Review and editing. This work was supported by the National Key R&D Program of China (2021YFA1202600), the Shanghai Sailing Program (23YF1402400), and China Postdoctoral Science Foundation (2022TQ0068). The authors declare no conflicts of interest.