A Wavelength Tunable Optical Neuron Based on a Fiber Laser
Yi Wei, Lili Gui, Fengbin Lin, Yihang Dan, Tian Yao Zhang, Xiaojuan Sun, Yueheng Lan, Kun Xu · 2021
With the increased applications of 5G communications, Internet of Things, autonomous driving etc., the world has been producing a large amount of data. Processing massive data intelligently with higher speed and lower energy consumption is hence extensively demanded. Photonic neuromorphic computing, which mimics the information processing manner of biological neural networks like human brains with optical components and systems, has attracted growing attention due to its great potential of meeting the aforementioned requirements. To this end, investigations and realizations of nonlinear optical systems that demonstrate similar spiking dynamics to biological neurons are of fundamental interests and importance. In this paper, we experimentally built a wavelength tunable fiber laser and numerically investigate its pulsing and nonlinear characteristics. The fiber laser exhibits excitability and behavesanalogously to a leaky integrate-and-fire (LIF) neuron. Such an optical neuron responds rapidly with maximum response frequency of about 17 kHz, at least 170 times faster than typical biological neurons. By adding a tunable optical bandpass filter (TOBPF) into the fiber laser cavity and adjusting its bandpass window, the operating central wavelength of the fiber laser can vary from 1530 nm to 1570 nm. With the spectral tunability, more possibilities and advantages in terms of computing performances by utilizing the wavelength degree of freedom could be envisioned, just like the case in optical communications where communication capacity can be easily enlarged with wavelength division multiplexing. We also numerically simulated the pulsing and nonlinear dynamics of the fiber laser with rate equations. In both experiment and simulation, the fiber laser shows the threshold behavior, temporal integration of nearby pulses, nonlinear transformation of input pulses, and features of refractory period, confirming its neuromorphic properties. With such a wavelength tunable neuromorphic neuron, we propose several applications. One is for typical logical digital computing such as AND and OR operation but with additional wavelength specificity; the second is for recognition of combined temporal and spectral patterns; the third is for spiking neural networks with wavelength multiplexing.