Harnessing Graded-like Spiking Dynamics in Semiconductor Lasers for High-Speed and Energy-Efficient Reservoir Computing

As machine learning scales, its computational and energy demands increase rapidly. Reservoir computing (RC), owing to its easy training and low hardware overhead, may offer viable solutions to the growing energy costs of machine learning. However, it faces significant bottlenecks in speed and energy efficiency when handling complex tasks. Inspired by biological vision, where graded neurons achieve high sensory precision with low energy, we develop a graded-like spiking reservoir architecture leveraging a semiconductor laser with controllable carrier dynamics. This hybrid approach is well-suited for high-speed and energy-efficient neuromorphic and photonic computing. By employing an electrical injection method, a solitary laser (implemented as one element of an integrated laser array) can demonstrate graded-like dynamics, bypassing the pulse rate limitations imposed by the refractory period or feedback loop, thereby enabling high-speed processing. The laser neuron operates without external perturbations or auxiliary components, forming a simple and energy-efficient core. Based on this, we construct an RC system that experimentally achieves 95.8% accuracy in MNIST digit classification and 91.8% in discrete-time bifurcation identification. Importantly, we numerically demonstrate that input encoding strategies can be seamlessly integrated into a graded-like spiking RC framework, significantly enhancing the computational performance without added hardware complexity. Furthermore, by incorporating the quasi-convolutional encoding algorithm, the normalized mean square error on the Mackey-Glass time series prediction task is experimentally reduced from 0.0114 to 0.0063. The numerical results show good qualitative agreement with the experiment. This work presents a generalizable and scalable framework for high-speed photonic neural computation.