All-analogue photonic artificial intelligence

Abstract

The article offers a methodological justification for developing a fully analogue photonic artificial intelligence (PAI) system based on optical technologies for processing continuous (analogue) signals without converting them into digital form. This makes it possible to overcome the limitations associated with digital data processing, namely the energy intensity and time-consuming nature of deep neural network training, the reduction of the natural signal spectrum, and the diminution of the weight of cognitive semantics of artificial intelligence (AI) models. The photonic approach makes it possible to reduce machine learning time and energy consumption significantly. It also helps to consider the non-formalizable aspects of the cognitive semantics of AI models to capture better human feelings, thoughts, and transcendent states of consciousness that digital computers cannot fully capture. Such advantages are achievable by replacing the multilayer structure of a neural network with a single matrix layer of Fourier convolution of a set of training images recorded on a holographic storage device. However, new problems arise on the photonic path of AI development, including synthesising unique photonic materials for rewritable three-dimensional holographic memory and controlling optical processes using digital computers. The article proposes an algorithm for the synthesis of photonic materials utilising an approach to solving inverse problems implemented using a genetic algorithm.