UCLA

In recent years, the perceived stagnation in the growth of electronic computing has fueled the search for alternative computing platforms. At the same time, the rise and success of deep neural networks have sparked renewed interest in neuromorphic computing for machine learning. Among these emerging platforms, optics stands out as a promising candidate for energy-efficient and low-latency computing. While much of the focus in optical computing has been on integrated optics, the recent success of diffractive deep neural networks or diffractive networks has also revived interest in diffraction-based free-space computing. Diffractive networks exploit programmed diffraction of light by spatially engineered surfaces for passive all-optical information processing at the speed of light propagation. These surfaces are digitally optimized through deep learning, enabling the input wave's diffraction to be tailored for a specific task by leveraging data that encapsulates the input-output relationship. Once the digital optimization is complete, these surfaces are fabricated and integrated into an all-optical ‘computer’, driven by passive light-matter interactions. Diffractive networks enable all-optical classification of input objects and support a diverse range of computational imaging tasks. This thesis presents methods to enhance the inference capabilities of diffractive networks and explores novel applications of this optical computing framework. First, the accuracy of diffractive systems for object classification is improved through ensemble learning, which involves combining the outputs of multiple diffractive networks. Novel strategies to diversify and assort the ensemble members are introduced, leading to significant performance gains. Then, time-lapse image classification is explored to improve the accuracy of standalone diffractive networks. By leveraging the natural jitter of the input object relative to the diffractive network, this method achieves, with a single diffractive network, accuracy comparable to that of an ensemble of 30 time-static networks. Next, the application of diffractive networks for all-optical hologram processing is demonstrated. This approach bypasses high-latency and power-intensive digital processing for removing twin-image artifacts, enabling passive inline hologram reconstruction at the speed of light propagation. Finally, a novel optical communication scheme is introduced, which utilizes electronic encoding and diffractive decoding to transfer images around arbitrarily shaped opaque occlusions of zero light transmittance. The joint training of the electronic encoder and the diffractive decoder allows efficient encoding of information in the transmitted wavefront, enabling the joint system to evade occlusions and transfer images even when the direct rays between the transmitter and the receiver are blocked. In summary, this thesis makes significant progress in the application of diffractive networks, while also laying the groundwork for future advancements in this field.