ECE PhD Prospectus Defense: Sina Moayed Baharlou

ECE PhD Prospectus Defense: Sina Moayed Baharlou

Title: End-to-End Learning for Image-Centric Optical Systems

Presenter: Sina Moayed Baharlou

Advisor: Professor Abdoulaye Ndao

Chair: Professor Kayhan Batmanghelich

Committee: Professor Abdoulaye Ndao, Professor Alexander Sergienko, Professor Lei Tian, Professor Kayhan Batmanghelich, Professor Jerome Mertz

Google Scholar Link: https://scholar.google.com/citations?user=e5okn4IAAAAJ&hl=en#

Abstract: Traditional optical systems follow a fixed pipeline in which optical elements modulate light and perform physical image formation, followed by a separate computational stage that reconstructs or interprets the captured measurements. This separation limits overall performance, since the optical front-end is not optimized jointly with downstream inference tasks. This research proposes a unified end-to-end deep learning framework for image-centric optical systems, where optical image formation is treated as a learnable, differentiable component within a machine learning pipeline. The key idea is to jointly optimize optical encoding and neural decoding so that sensing and inference are co-designed for task-driven performance. By embedding optical models into deep learning frameworks, the approach enables data-driven solutions to inverse problems in imaging, wavefront sensing, and optical device design. This research addresses key challenges in full-stack end-to-end optical design, including ill-posed inverse problems, by developing efficient deep learning-based inverse design pipelines and computational methods that are both fast and memory-efficient for practical use. These components are aimed at supporting the physical realization of intelligent optical systems. Overall, the work advances end-to-end machine learning for optics by bringing together physical sensing, computational reconstruction, and device design within a unified differentiable learning framework.