UCLA

The remarkable success of physics in explaining nature and engineering machines is rooted in low-dimensional deterministic models that accurately describe a wide range of natural phenomena. In contrast, Artificial Intelligence (AI) led by neural networks has introduced an alternate data-driven computational framework, achieving astonishing performance in various fields. However, this impressive feat comes at the cost of challenges, including the need for massive datasets, high computational cost, and limited interpretability. To address these issues, there is a growing interest in hybrid approaches that merge the laws of physics and physical systems with AI. This dissertation focuses on developing efficient and interpretable algorithms for vision enhancement and AI applications by drawing inspiration from the principles of optical physics. Unlike traditional algorithms that rely on hand-crafted empirical rules or neural networks that are data-driven and computationally heavy, physics-inspired algorithms leverage laws of nature as blueprints, resulting in low dimensionality, high efficiency, and full interpretability.

The dissertation begins with the principles of Photonic Time Stretch, a hardware technique grounded in optical physics for ultrafast and single-shot data acquisition, which seeded the ideas herein. A unified physical and mathematical framework is derived based on the governing Nonlinear Schrödinger Equation (NLSE). Inspired by this physical principle, a novel class of algorithms called PhyCV (Physics-inspired Computer Vision) is developed for image and video processing tasks, including edge detection, low-light enhancement, and motion detection. These algorithms demonstrate superior performance in vision enhancement and are well-suited for applications on mobile and edge devices with constrained resources. Furthermore, a novel physics-based AI model called the "Nonlinear Schrödinger Network" is proposed, which treats the NLSE as a trainable model in the numerical domain to learn complex patterns and nonlinear mappings from data. This physics-inspired approach provides a more interpretable and parameter-efficient alternative compared to black-box neural networks, achieving comparable or better accuracy on nonlinear classification tasks while significantly reducing the required number of parameters.