Fig. 1
From: Optical neural networks with intensity‑based projection layers as effective nonlinear activations
Schematic of the complex-valued neural network (CVNN) with projection layers and the experimental implementation setup. (a) Schematic of a traditional RVNN utilizing predefined nonlinear activations, consisting of two iteration blocks (IBs). (b) Schematic of the CVNN with projection layers, also comprising two IBs. The input can be encoded as either amplitude or phase profile of the electric field. By applying the complex weight matrix, the complex-valued output is generated and subsequently projected onto the intensity. This projection functions, together with encoding, as a nonlinear activation mechanism analogous to that in traditional RVNNs. (c) Experimental setup for implementing the optical CVNN with projection layers. The input is generated using a digital micromirror device (DMD) or spatial light modulator (SLM) for amplitude or phase encoding, respectively. The second SLM implements the complex weight matrix. The intensity of the electric field is measured by a camera, which can be used as the input for the subsequent IB through cascading.
