Quantized Optical Neural Network Based on Microring Resonators with On-Chip Modulation

Yun Hu, Huifan Zhang, Pingqiang Zhou

Abstract: Optical computing has the advantages of low energy consumption, high speed, and parallelism. It is well suited to the high-performance implementation of neural network. As the core units for optical computing, microring resonators (MRRs) still face several challenges. First, large scale MRR arrays are difficult to implement, which limits the integration of neural networks on optical chips. Second, the wavelength sensitivity of MRR would result in higher cumulative errors in highprecision neural networks. To solve these problems, in this paper, we propose an on-chip modulated MRR array to improve the chip integration. On-chip modulated MRR arrays realized more tightly integrated optical chips. Moreover, we adopt the optimized partial Euler bending MRRs to reduce the transmission loss through the more smaller radius waveguide bending. To reduce cumulative errors, based on our MRR architecture, we propose a quantization simulation design on the optical neural network (ONN). The quantized optical neural network helps reduce the impact of systematic errors. It also can optimize the computing performance of the chip. The quantization experimental results on different bit widths and different channel widths provide a demonstration of the advantages of our integration of MRRbased optical computing and the development of ONN schemes.