Fig. 1: FFM onsite machine learning for optical systems.
From: Fully forward mode training for optical neural networks
a, Conventionally, optics-related AI is designed through offline modelling and optimization, leading to limited design efficiency and system performance. b, General optical systems, including free-space lens optics and integrated photonics, contain the modulation regions (dark green) and propagation regions (light green), where the refractive indexes are respectively tunable and fixed. c, These regions in the optical system can be mapped to weights and neuron connections in the neural representation, which enables the construction of a differentiable onsite neural network between the input and output (top-left panel). With spatially symmetrical reciprocity, the data and error computes share forward physical propagations and measurements, and the onsite gradients are calculated for the update of the refractive indexes in the design region (top-right and bottom-left panels). Through onsite gradient descent, the optical system gradually converges (bottom-right panel). RI, refractive index; A, amplitude; Φ, phase.
