Learning Neural Representations of Phase Relationships in Imaging Problems

Abstract: Machine learning has brought significant advancements in capabilities for social media platforms and entertainment services, and is increasingly becoming part of our everyday experience. On the other hand, this same level of adoption has not been observed in scientific imaging problems underlying adaptive optics, free space optical communication, or crystallography. The fundamental reason for this gap is likely due to the black box nature of machine learning. Instead of aiming to make machine learning more transparent, can we instead go to the physics equations, arrive at an important and narrow yet analytically intractable task, and let a small network bridge this gap for us? In this talk I will discuss some recent advancements in the problems of imaging through turbulence and wavefront estimation that use neural representations to model the connection between phase and space in imaging problems. I will also outline how similar representations can be extended to a broader class of imaging problems while respecting the underlying physics.

Bio: Nick Chimitt is currently a research scientist at Purdue University. He received his PhD from Purdue in 2023. He is the co-inventor of the Zernike-based propagation-free atmospheric turbulence simulator and the phase-to-space transform. He is a co-author of the book Computational Imaging through Atmospheric Turbulence (Now Publisher, 2023) and has delivered tutorials at various conferences including CVPR and ICIP. He was also the organizer of the UG2+ workshop at CVPR 2024. His research interests include imaging through turbulence, phase-based optical modeling, machine learning, phase retrieval, and computational imaging.

Learning Neural Representations of Phase Relationships in Imaging Problems

Abstract: Machine learning has brought significant advancements in capabilities for social media platforms and entertainment services, and is increasingly becoming part of our everyday experience. On the other hand, this same level of adoption has not been observed in scientific imaging problems underlying adaptive optics, free space optical communication, or crystallography. The fundamental reason for this gap is likely due to the black box nature of machine learning. Instead of aiming to make machine learning more transparent, can we instead go to the physics equations, arrive at an important and narrow yet analytically intractable task, and let a small network bridge this gap for us? In this talk I will discuss some recent advancements in the problems of imaging through turbulence and wavefront estimation that use neural representations to model the connection between phase and space in imaging problems. I will also outline how similar representations can be extended to a broader class of imaging problems while respecting the underlying physics.

Bio: Nick Chimitt is currently a research scientist at Purdue University. He received his PhD from Purdue in 2023. He is the co-inventor of the Zernike-based propagation-free atmospheric turbulence simulator and the phase-to-space transform. He is a co-author of the book Computational Imaging through Atmospheric Turbulence (Now Publisher, 2023) and has delivered tutorials at various conferences including CVPR and ICIP. He was also the organizer of the UG2+ workshop at CVPR 2024. His research interests include imaging through turbulence, phase-based optical modeling, machine learning, phase retrieval, and computational imaging.