Physical Layer-Aware Digital-Analog Co-Design for Photonic Convolution Neural Network

Abstract

Photonic convolution neural network (CNN) has shown tremendous capability to provide Tera-level operations/sec (OPS) for streaming object detection with ultra-low energy consumption. However, due to the nature of analog photonic computing which is constrained by accumulated noise from various sources, its physical operating precision is significantly lower than that of electronic artificial intelligence (AI) accelerators powered by tensor processing unit (TPU) and graphic processing unit (GPU). In this article, we propose a robust digital-analog co-designed photonic CNN using error learning and Kirsch edge enhancement. Experimental results show that, nearly 96.1% recognition accuracy, compared with 34.2% in baseline photonic CNN, can be achieved at global 2-bit precision through physical error learning in dense layer and edge enhancement by photonic Kirsch operator. The proposed photonic CNN architecture is promising to provide 10 Tera-OPS scale computing speed deployed by the state of art PAM-4 optical modules with high energy efficiency.