UCLA

Optical computing and information processing, known for its energy efficiency, scalability, low latency, and parallelism, stands at the forefront of revolutionizing computational systems and machine vision. Among different optical computing designs, diffractive optical neural networks represent a free-space-based framework that can be used to perform computation, statistical inference, and inverse design of optical elements. A diffractive neural network is composed of multiple transmissive and/or reflective diffractive layers (or surfaces), which leverage light-matter interactions to jointly perform modulation of the input light field to generate the desired output field. These passive diffractive layers, each containing numerous spatially engineered diffractive features, are optimized in a computer using deep-learning tools. Post-training, these layers form a passive optical processing unit powered solely by illumination light, capable of processing based on 2D/3D input information at the speed of light and analyzing various light properties e.g., amplitude, phase and polarization. Building upon these foundations, this thesis presents multiple strategies and designs to enhance the performance and capacity of diffractive networks for all-optical information processing. First, a differential detection scheme is introduced to improve the classification performance by addressing the non-negativity constraint in opto-electronic detectors. Additionally, the parallel training of multiple networks, each targeting different object classes or trained diversely, further boosts the overall inference performance. Extending the framework to harnessing broadband light, the encoding of spatial information into the power spectrum of diffracted light is presented, facilitating classification and reconstruction with a single-pixel spectroscopic detector. Furthermore, diffractive networks have been theoretically and numerically proven as a universal linear transformer, capable of all-optically implementing arbitrary complex-valued linear transformations between its input and output field-of-views. Introducing the wavelength multiplexing scheme into a broadband diffractive processor allows simultaneous execution of a large group of linear transformations across different wavelength channels, thereby enhancing the computational throughput of the framework. Another significant advancement is made with the integration of polarization-sensitive elements into isotropic diffractive networks, which enables universal polarization transformations of spatially varying polarization fields, surpassing the limitations of traditional polarization control methods. In summary, this thesis significantly advances diffractive neural networks as a versatile all-optical information processing platform that is ideal for efficient, high-speed and intelligent machine vision systems, also laying the groundwork for future innovations.