Abstract

With the benefit of high speed, wide bandwidth, and low latency, optical neural networks (ONN) form a promising neuromorphic computing platform for accelerating artificial intelligence tasks. ONNs contain two kinds of fundamental functional operations: multiply-accumulate (MAC) stages and nonlinear activation function units, as shown in Fig.1(a). In the past years, different optical MAC units have been developed using wavelength-division multiplexing (WDM) schemes [3]. However, the optical nonlinear activation function units still remain a challenge for the integration of large-scale ONNs. Most nonlinear activation functions are implemented in optical-electrical (OE) conversion schemes [2-3], where processing speed is limited by the latency time of OE conversion. Therefore, an integrated all-optical nonlinear activation function is an attractive solution for optical neuromorphic computing.

© 2023 IEEE