Engineering nonlinear activation functions for all-optical neural networks via quantum interference (2504.04009v1)

Abstract: All-optical neural networks (AONNs) harness the wave nature of light to achieve unparalleled speed and energy efficiency for artificial intelligence tasks, outperforming their electronic counterparts. Despite their potential, the development of deep AONNs is constrained by the high optical power demands of conventional nonlinear optical processes, which limits scalability. This work introduces a novel low-power nonlinear optical activation function scheme based on a three-level quantum medium driven by two laser fields. Unlike traditional single-input, single-output activations, our design offers two-port optical nonlinear activation functions with both self- and cross-nonlinearities, making them great candidates for multi-input, multi-output networks. The approach allows precise control of nonlinear optical behavior, achieving sigmoid and rectified linear unit (ReLU)functions at ultralow power levels (~ 17 uW per neuron). Our theoretical and numerical analysis demonstrates the feasibility of constructing large-scale, deep AONNs with millions of neurons powered by less than 100 W of optical power. This advancement represents a significant step toward scalable, high-speed, and energy-efficient AONNs for next-generation AI hardware.