Emergence of intelligence with complexity in integrated optical system
Photonic integrated circuits are always wired as an optical analogue of FPGAs and ASIC. In this article, we discard the framework of electronic circuits, and explore an new way of building photonic processors by utilizing the wave nature of light.
As the depth of a machine learning system outweighs the number of elements per layer, here we demonstrate high accuracy image classifications at telecommunication wavelengths in the multi-layer one-dimensional metasurface systems, with CMOS foundry compatible process. Arrays of rectangular slots are defined in the silicon layer. The slot lengths in those phase-only transmissive arrays are pre-trained by deep diffractive neuron networks. Beyond conventional classification functions, the metasystems also demonstrate unique functions of wavelength demultiplexing and multi-wavelength pattern classifications, with potential applications from spatial division multiplexing based optical interconnects to machine vision.
The new way of creating photonic integrated circuits (PIC) with cascaded metasurface layers, named metasystem, offers the possibility of integration thousands of programmed phase shifters in a sub mm area. The manifested integration scale brings the intelligence to the PIC - which can handle errors and uncertainties in the inputs, and perform the 'hyperspecial image classifications'.
Fig. 1: Integrated metasurface system for spatial pattern classification.