UCLA

Deep learning has been revolutionizing information processing in many fields of science and engineering owing to the massively growing amounts of data and the advances in deep neural network architectures. As these neural networks are expanding their capabilities towards achieving state-of-the-art solutions for demanding statistical inference tasks in various applications, there appears to be a global need for low-power, scalable and fast computing hardware beyond what existing electronic systems can offer. Optical computing might potentially address some of these demands with its inherent parallelism, power efficiency, and high speed. Recent advances in optical materials, fabrication, and optimization techniques have significantly enriched the design capabilities in optics and photonics, leading to various successful demonstrations of guided-wave and free-space computing hardware for accelerating machine learning tasks using light. While integrated waveguide-based photonic approaches mainly aims to replace the current electronic computing hardware with better alternatives, free-space optical neural network architectures and related computing techniques offer unique advantages particularly for inference tasks in visual computing applications, where the information is already in the optical domain.This dissertation introduces diffractive optical networks that are designed based on Diffractive Deep Neural Networks (D2NN) framework using deep learning to tackle various challenges in computational machine vision by providing power-efficient, fast, scalable and massively parallel all-optical solutions. First, a series of design advances were devised to improve the statistical inference accuracy of diffractive object classifiers. Second, hybrid (optical-electronic) neural network systems, which uses diffractive optical networks as front-end optical processors preceding back-end electronic neural networks, were investigated to enable task-specific camera systems that can perform object classification with fewer pixels, thus with less memory and power consumption. In addition, D2NN framework was extended to mitigate the adverse impact of possible physical error sources, termed as vaccinated-D2NN (v-D2NN). The success of v-D2NN was experimentally demonstrated at THz wavelengths by comparing the classification accuracies of 3D-printed nonvaccinated and vaccinated diffractive handwritten digit classifiers under the presence of layer-to-layer misalignments. Next, a diffractive all-optical object classifier was designed to provide inference accuracy that is invariant under random changes on the scale, position and orientation of the input objects with respect to the diffractive surfaces. Furthermore, the all-optical information processing capacity of diffractive optical networks was studied to prove that the dimensionality of the solution space representing the set of all-optical transformations established by a diffractive network increases linearly with the number of diffractive surfaces, up to a limit determined by the size of the input/output fields-of-view. In parallel, the diffractive optical networks were shown to all-optically perform arbitrary complex-valued linear transformations, including space-variant operations, noninvertible and nonunitary matrices, with negligibly small errors provided that the total number of diffractive neurons is sufficiently large to satisfy space-bandwidth product demands on input and output fields-of-view. A diffractive permutation network that can all-optically implement 625 interconnects between its input and output was fabricated using 3D printing and its performance was demonstrated at THz wavelengths. Beyond the outlined optical computing and machine learning applications, diffractive optical networks can also be utilized to all-optically solve challenging inverse problems in computational imaging. Highlighting this aspect, diffractive optical networks that can all-optically perform phase retrieval to reveal the quantitative phase image (QPI) of weakly scattering objects were devised. Based on the conducted analysis, these diffractive QPI networks can resolve subwavelength features, ~0.67λ, of an input phase object, with λ denoting the wavelength of illumination. Finally, in certain application scenarios, spatial overlap between phase objects poses an irreversible information loss due to the superposition of individual phase delays. It was demonstrated that diffractive optical networks can be trained to solve this challenging problem to infer the classes of spatially overlapping phase objects. Moreover, when these diffractive phase object classifiers are combined with electronic deep neural networks, the individual phase images of the objects spatially overlapping within the input field-of-view can be recovered based on the all-optically synthesized class scores, despite the phase ambiguity. All the studies presented in this dissertation demonstrating the success of diffractive optical networks in various general-purpose computing, statistical inference and inverse computational imaging tasks can potentially lead them to largely replace conventional optical components in the next-generation, task-specific machine vision designs that can achieve a given task with fewer pixels, leading to faster, more memory- and power-efficient systems.