Photonic Integrated Reconfigurable Linear Processors as Neural Network Accelerators

Reconfigurable linear optical processors can be used to perform transformations and are instrumental in effectively computing matrix–vector multiplications required each neural network layer. In this paper, we characterize compare two thermally tuned photonic integrated realized silicon-on-insulator silicon nitride platforms suited for extracting feature maps convolutional networks. The reduction bit resolution when crossing the processor is mainly due losses, range 2.3–3.3 chip 1.3–2.4 chip. However, lower extinction ratio of Mach–Zehnder elements latter platform limits their expressivity (i.e., capacity implement any transformation) 75%, compared 97% former. Finally, outperforms one terms footprint energy efficiency.