Now, writing in Nature, Dong, Brückerhoff-Plückelmann and colleagues demonstrate that low-coherence light can, somewhat counterintuitively, be employed for photonic convolutional computing, and even enhance it (B. Dong et al. Nature 632, 55–62; 2024).

The low-coherence light, produced by filtering the emission from an erbium-doped fibre amplifier, is evenly split by amplitude between N input channels, with each channel's amplitude modulated to form the N-element input vector (see picture). This input vector is then sent into a chip with weighting elements to perform matrix calculations. On the other hand, a coherent system requires wavelength multiplexing, demanding N separate optical bands to achieve the same N-element vector while avoiding unwanted interference. Thus, the incoherent approach increases parallelization, as N times more operations can be performed simultaneously within a single frequency band compared to a coherent system.

现在，Dong、Brückerhoff-Plückelmann 及其同事在 Nature 上撰文证明，低相干光可以用于光子卷积计算，甚至增强它（B. Dong et al. Nature632， 55–62; 2024）。

通过过滤掺铒光纤放大器的发射产生的低相干光，在 N 个输入通道之间按振幅均匀分配，每个通道的振幅被调制以形成 N 晶片输入矢量（见图）。然后将此输入向量发送到带有加权元素的芯片中，以执行矩阵计算。另一方面，相干系统需要波长多路复用，需要 N 个单独的光带来实现相同的 N 元素矢量，同时避免不必要的干扰。因此，非相干方法提高了并行化，因为与相干系统相比，在单个频带内可以同时执行 N 倍的操作。