Optical computing with spatiotemporal fiber nonlinearities

Optics offers significant advantages for information processing and artificial intelligence in terms of information density, speed, and energy efficiency compared to electronics. Starting with the Hopfield model’s implementation, the initial optical computing studies focused on mimicking trainable artificial neural network (ANN) architectures [1] . Including memory storage methods with holography, various methods were proposed, yet the controllability level of the training optical networks still challenging [2] . In the last decade, brain-inspired ANN designs operating based on projection of information to high order dimension with a fixed parameter ANN structure followed by a trainable decision layer were introduced [3] , [4] . The fixed network structure makes fixed-parameter ANN designs easy to implement with physical systems and various optical implementations were demonstrated [5] . However, optical nonlinear activation implementation was challenging to realize, and generally digital signal processing tools are employed [6] .