Photonic-digital hybrid artificial intelligence hardware architectures: at the interface of the real and virtual worlds

Abstract

Artificial intelligence (AI) technology has become an undeniable presence in society, as seen in the growing use of chatbots (e.g. ChatGPT) and recent Nobel Prize awards. However, the challenge of developing AI hardware architectures that efficiently integrate real-world analog signals with digital computational frameworks remains unresolved. This is particularly true for optical and neuromorphic computing systems. This work reports the implementation of an optical computing module in two hybrid AI architectures. The key element of the module is a photonic layer comprised of nanometer-scale spherical carbon dots. When excited with light-emitting diodes, the photonic layer exhibits a complex optical response. This layer is a feasible building block for two hybrid AI architectures: one based on optical processing and the other on the principles of neuromorphic computing. The modality of operation of the photonic layer is that it converts numerical input into a complex emission space, the results of which are processed using a Gaussian process model. Two systems based on this building block have been tested on a real-world dataset. They outperform conventional digital-only models, with coefficients of determination of r2 = 0.90 and r2 = 0.84 (training) and r2 = 0.85 and r2 = 0.87 (testing) for the optical and neuromorphic architectures, respectively. The systems discussed in this work serve as interfaces that bridge the real and virtual worlds. They offer exceptional optical properties, relative innocuity, low toxicity, stability, cost-effectiveness, seamless scalability, and robustness. The proposed architectures push the boundaries of advanced hybrid designs.