Recently, Prof. Hao Yue and Prof. Shuiying Xiang’s research group at Xidian University has made important progress in the research field of photonic neuromorphic computing. They published an article in Optica entitled “Hardware-algorithm collaborative computing with photonic spiking neuron chip based on integrated Fabry–Perot laser with saturable absorber”.
The human brain has the inherent advantages of low power consumption, high robustness, efficient parallelism and self-adaptation. Neurons and synapses are two key fundamental unitswithin the human brain. The neuromorphic computing, which is inspired by the human brain, is a new competitive computing paradigm that seeks to overcome the Von-Neumann bottleneck in the post-Moore’s era. In neuromorphic computing systems, the computing hardware and algorithmsare developed based on the structure and information processing mechanism of biological neural networks. Although electronic neuromorphic computing chips have made great progress, they are still limited in speed and power consumption due to electronic bottlenecks and the slowing down of Moore's Law. Withsuch remarkable advantages as ultra-high speed, large bandwidth and multiple dimensions, photonic neuromorphic computinghas the potential to overcome the limitations of electronic neuromorphic computing. Linear and nonlinear computing units are indispensable fundamental units for integrated photonics neuromorphic computing systems. The photonic implementation of linear computation has been well developed, but nonlinear computation is still the most challenging task of photonic neural networks.
Forthis nonlinear computing challengein photonic neural networks, a new photonic spiking neuron chip was proposed and fabricated based on a Fabry–Perot laser with a saturable absorber. The controllable neuron-like nonlinear response (including temporal integration, excitability threshold,inhibitory dynamics and refractory period) was verified by experiments, where the spike processing speed canreach as great as 10GHz, which is 7 orders of magnitude faster than the response rate of biological neurons, and the energy consumption is about 7.3fJ/spike. In order to avoid the limitation of hardware integration scale, a time-division multiplexedspatialtemporal coding mechanism was proposed in the optical domain, which greatly reduces the requirement for hardware nodes. To the best of our knowledge, the hardware and algorithm collaborative computing of photonic spiking neural networksis experimentally demonstrated for the first time. The architecture is shown in Figure 1, and the spike-based pattern recognition task isshown in Figure 2. This experimental finding is an important step to promote the practical application of integrated photonic spiking neural network chips. It also proves the potential of constructing large-scale multi-layer photonic spiking neural network chips to solve complex tasks, laying an important foundation for the hardware implementation of large-scale photonic spiking neural networks.
Figure 1. Hardware-software cooperative computing architecture of photonic spiking neuron chip based on a spatiotemporal encoding mechanism
Figure 2. Pattern recognition taskimplementation based on the hardware-software cooperative computing architecture of the photonic spiking neuron
The photonic spiking neuron chip reported in this article was fabricated based on the traditional InP-based laser technology platform. It has the characteristics of easy integration, low power consumption, high speed and easy tuning. It is suitable for application scenarios requiringa large bandwidth, high speed and low latency, and lays the device foundation for realizing integrated photonic neuromorphic computing systems. It is expected to be uniquelycompetitive in such applications as data center, edge computing, autonomous driving,etc.
The research has been carried out under theextensive collaboration between the photonic neuromorphic computingteamatXidian University and the team of Prof. Xiangfei Chen and Associate Professor Yuechun Shi from Nanjing University.Xidian University is the leading research unit. Prof. Shuiying Xiang and Associate Prof. Yuechun Shi are the co-first authorswith equal contributions, and Prof. Xiang is the corresponding author. The other cooperating organizations include the Institute of Semiconductors inChinese Academy of Sciences, University of Glasgow, and Nantong University. This work was supported by the National Key Research and Development Program of China(2021YFB2801900), the National Outstanding Youth Science Fund Project of National Natural Science Foundation of China (62022062) and the National Natural Science Foundation of China (No. 61974177, No.61674119).
Other related works
In recent years, Prof. Xiang’s group has focused on research on the chip and algorithm of photonic neuromorphic computing. The basic theory, fundamental device and key technology, and integrated chip and core algorithm have been studied systematically. From the perspective of photonic neuromorphic devices, the nonlinear neuron-like response of photonic spiking neurons based on discrete optical devices[J. Lightwave Technol. 36(19), 4227, 2018; Opt. Lett., 44(7):1548-1551, 2019] and the synaptic plasticity property of the photonic synapse were demonstrated numerically and experimentally[Sci China Inf Sci, 65(8): 182401, 2022]. Moreover, the all-optical XOR was numerically and experimentally demonstrated based on a single spiking neuron [Photonics Res., 9(6):1055, 2021; Opt. Lett., 45(5), 1104, 2020]. In addition, the photonic spiking neuron was also demonstrated to perform all-optical binary convolution and image edge detection [Photonics Res., 9(5):201, 2021]. From the perspective of core algorithm, the self-consistent theoretical model was developed, and the unsupervised/supervised training algorithm was designed[IEEE Trans. Neural Networks Learn. Syst., 32(6), 2494, 2021;IEEE J. Sel. Top. Quantum Electron., 25(6): 1700109, 2019]. In addition, the supervised training algorithm that incorporated weight-delay plasticity was further proposed [Photonics Res., 9(4):119, 2021], and the supervised training algorithm for the multi-layer photonic spiking neural network was developed to realize the pattern recognition task. From the perspective of theintegrated chip, the prototypical photonic neuromorphic computing system based on the fabricated photonic spiking neuron was experimentally demonstrated [Photonics Res., 11(1), 65, 2023]. These research results provide an important theoretical and device basis for the development of photonic neuromorphic computingintegrated chips.
Original article link: https://doi.org/10.1364/OPTICA.468347