Fig. 2: Light field neural network (LFNN).
From: Optical neural network via loose neuron array and functional learning
The LFNN prototype consists of an input plane, an output plane, two liquid-crystal layers, and three perpendicular linear polarizers. The output plane is a scattering plane followed by a camera to acquire the data. We use an LCD as the input plane by representing artificial neurons with pixels. Like a fully connected neural network layer, these neurons of the input plane synergistically activate the artificial neurons of the next layer through light propagation. We adopt two LCDs as controllable liquid-crystal components by eliminating their backlights and polarizing films. The key idea of the LFNN is to train the pixel control parameters of the liquid-crystal components to distort the polarization directions of the input light so as to synthesize the attenuated light paths as a weighted connection between the input and output planes. In addition to the parameters of LCs, we modify the driver so that the neuron-wise intensity gains of the input and output planes are trainable parameters to increase the degree of freedom. The resolutions of the input, output, and LCs are all 32 × 32 with RGB channels in our prototype.
