Programmable Temporal Dynamics in a Laser Neuron for High‐Speed Photonic Reservoir Computing

Abstract

ABSTRACT Memory and nonlinearity play crucial roles in reservoir computing (RC) performance, with different tasks requiring different levels of these properties. However, in most physical RC systems, memory and nonlinearity are fixed or can only be changed through hardware redesign, significantly limiting the system's adaptability. Here, an adaptive photonic RC system is experimentally demonstrated based on a high‐speed laser graded neuron with programmable dynamics, encompassing both memory and nonlinearity. By simply tuning the current and voltage biases of the laser neuron within a small range ( mA, V), both properties can be simultaneously programmed, enabling adaptive optimization for applications with diverse memory and nonlinearity requirements. The proposed system demonstrates effective sequence prediction performance, with dynamics programming reducing prediction errors in multiple tasks, and successfully demonstrates proof‐of‐concept image generation. Specifically, optimizing laser dynamics reduces error across the tested bias states in a sequence prediction task, and increases the structural similarity index from 0.573 to 0.705 for image generation using a single laser neuron. Furthermore, a multi‐dynamics joint training approach is proposed to further boost system performance. This work lays the foundation for adaptive and versatile RC systems, paving the way for high‐speed, intelligent signal processing.