搜索

x
中国物理学会期刊

半导体环形激光器中基于多模时序编码的无反馈光子储备池计算

Feedback-free photonic reservoir computing based on multimode temporal coding in semiconductor ring lasers

PDF
导出引用
  • 半导体环形激光器 (SRL) 天然支持顺时针与逆时针两个反向传播模式, 是构建光子神经形态计算的重要器件。然而, 基于 SRL 的光子储备池计算大多依赖物理反馈环来提供衰落记忆, 存在参数整定困难、稳定区间有限等问题。本文将单个 SRL 与受生物启发的准卷积 (QC) 编码相结合, 提出一种无反馈光子储备池计算方案, 依据 QC 编码施加位置的不同, 构建了输入编码型 (IC-RC)、输出编码型 (OC-RC) 与混合编码型 (HC-RC) 三种互补储备池变体, 三者分别通过在输入端、输出端或两端同时施加编码来提供差异化的算法记忆。通过建立速率方程模型, 在混沌时间序列预测、非线性信道均衡与线性记忆容量三项基准任务上, 分析了注入强度、频率失谐以及 QC核关键参数对三种方案性能的影响。结果表明, 三种方案均明显优于传统时延储备池计算。本文还发现不同的编码方案具有不同的优势, 其中 IC-RC 预测性能最优, HC-RC 记忆保持能力最强,输出层激活函数可进一步改善预测与信道均衡性能, 而对记忆容量影响甚微。本文结果不仅验证了所提方案的有效性, 也为构建高度可集成的光子神经形态计算系统提供了新的思路。

     

    Semiconductor ring lasers (SRLs) naturally support clockwise (CW) and counterclockwise (CCW) counter-propagating modes, and thus offer intrinsic dual-channel parallelism in a compact cavity well suited to photonic integration. However, photonic reservoir computing (RC) built on such lasers usually relies on an external optical feedback loop to provide fading memory, which makes the operating parameters difficult to calibrate and confines the system to a narrow stable operating range. In this work, a feedback-free photonic RC scheme is proposed by combining a single SRL biased in its bidirectional emission regime with bio-inspired quasi-convolution (QC) coding, in which a triangular kernel with linearly decreasing weights algorithmically embeds tunable short-term memory into the data stream. The kernel size Q and the step coefficient \beta independently control the depth and the temporal granularity of this memory. According to where the kernel is inserted, three complementary variants are constructed, an input-coding variant (IC-RC), in which the CW mode is driven by the uncoded masked signal and the CCW mode by the coded signal, so that the two modes acquire differentiated temporal contexts, an output-coding variant (OC-RC), in which the kernel instead acts on the detected mode intensities, and a hybrid-coding variant (HC-RC), in which independent kernels are applied at both the input and the output. The two mode intensities are sampled and concatenated into a single readout matrix, which doubles the number of virtual nodes without additional hardware. Based on a rate-equation model, the three schemes are benchmarked against a conventional time-delay reservoir computer (TDRC) on Santa Fe chaotic time-series prediction, nonlinear channel equalization, and linear memory capacity (MC). All three variants outperform the TDRC and maintain low, plateau-like errors over the entire injection-strength scan, broadening the high-performance region in the plane of injection strength and frequency detuning by more than an order of magnitude. The two families of tasks are found to be governed by different kernel properties, prediction and equalization depend on the overall richness of the temporal context and therefore vary smoothly with the kernel parameters, whereas the memory capacity depends on the precise alignment between the kernel delay and the sampling grid and thus rises in discrete steps. IC-RC achieves the highest prediction accuracy and HC-RC the largest memory capacity, while the output-layer activation function further improves prediction and equalization but has little effect on the MC. These results provide a theoretical basis for feedback-free photonic RC and offer new insight into the development of highly integrable photonic neuromorphic computing systems.

     

    目录

    /

    返回文章
    返回