HNLF-based feedforward photonic reservoir computing with adaptive memory

Abstract

Time-delayed reservoir computing has emerged as an energy-efficient neuromorphic paradigm due to its minimal physical footprint, yet its reliance on feedback-loop-induced fading memory fundamentally limits scalability and task adaptability. In this work, a novel photonic feedforward architecture that eliminates feedback mechanisms while enabling memory-tunable operation is proposed. By exploiting the intrinsic nonlinear response and the light dispersion characteristic of highly nonlinear fiber (HNLF), the system achieves baseline memory properties suitable for simple temporal tasks. For enhanced memory-intensive processing, a dynamic input encoding scheme that systematically modulates temporal correlations without physical structural modifications is utilized. The passive low-loss HNLF implementation ensures ultralow power consumption and broad operational bandwidth, overcoming the bandwidth constraints of active feedback components. The feasibility and effectiveness of the proposed architecture are experimentally validated on two benchmark tasks with distinct memory requirements, the Santa Fe chaotic time series prediction and the NARMA10 prediction. Competitive performance is achieved, with normalized mean square errors of 0.0049 and 0.2159 for the Santa Fe and NARMA10 tasks, respectively.