Because of the von Neumann bottleneck, neuromorphic networks aimed at in-memory computing, such as brains, are extensively studied. As artificial synapses are essential in neuromorphic networks, a photonic synapse based on slot-ridge waveguides with nonvolatile phase-change materials (PCMs) was proposed and demonstrated in an SOI platform with standard complementary metal-oxide-semiconductor (CMOS) process for a larger weight dynamic range. The change of the optical transmission spectrum of our photonic synapses was about 3.5dB higher than that of primitive synapses, which meant large weight dynamic range and more weight values. A 90.7% recognition accuracy based on our photonic synapses, which was 2.6% higher than that of primitive synapses, was realized for the MNIST handwritten digits recognition task performed by a three-layer perceptron. Besides, because of the nonvolatile nature of PCMs, the weights achieved by our photonic synapses can be stored in situ ensuring a lower consumption in in-memory computing. This framework can potentially achieve a more efficient in-memory computing neuromorphic network in silicon photonics.

中文翻译：

---

基于槽脊波导和PCM的片上光子突触，用于内存中计算

由于冯·诺依曼（von Neumann）瓶颈，针对内存计算（例如大脑）的神经形态网络已得到广泛研究。由于人工突触在神经形态网络中必不可少，因此提出了一种基于带有非易失性相变材料（PCM）的缝脊波导的光子突触，并在具有标准互补金属氧化物半导体（CMOS）工艺的SOI平台中进行了演示。体重动态范围。我们的光子突触的光传输谱的变化比原始突触的光传输谱的变化高约3.5dB，这意味着较大的重量动态范围和更大的重量值。通过三层感知器执行的MNIST手写数字识别任务，基于我们的光子突触的识别精度达到90.7％，比原始突触的识别精度高2.6％。此外，由于PCM的非易失性，我们的光子突触所获得的权重可以就地存储，从而确保了内存计算的较低消耗。该框架可以潜在地在硅光子学中实现更有效的内存中计算神经形态网络。