CSE Distinguished Seminar | George Barbastathis

Abstract

For the past decade, my group has been working on machine learning for computational imaging and, more specifically, supervised training of regularizers for ill-posed and ill-conditioned inverse problems. The notion of data-driven regularization has been around since at least the invention of learned sparse codes, or “dictionaries,” by Olshausen and Field in 1996. Learning the code through a deep neural network, as pointed out in 2010 by Gregor and Lecun, narrows the regularizer on-demand. Thus, improvements should be expected in terms of resilience to noise and incomplete measurements. Indeed, circa 2017~18, Unser’s group at EPFL devised a cascaded neural proximal gradient scheme for linear tomography with reduced exposure dose; Ozcan’s group at UCLA showed that a low-NA microscope objective can capture features of comparable fidelity as high-NA; and my group at MIT discovered that a Gerchberg-Saxton-Fienup  module followed by a deep neural network can retrieve the phase of the complex field robustly from highly noisy intensity measurements.

Using neural networks for imaging in various contexts is now widespread. What have we learned from these deployments? It is certainly true that neural networks capture priors effectively. It is less clear how optical physics, to the degree that it is involved in machine training and if at all, influences the reconstruction accuracy. I will discuss this topic in the context of more recent work by my group on image-based parameter estimation for dynamical phenomena that are severely undersampled in both space and time as well as severely photon-limited.

Acknowledgment: Recent work was funded by the US Air Force Office of Scientific Research, Takeda Pharmaceuticals, and Fujikura Ltd. The author acknowledges working with Claude, an artificial intelligence; the author has reviewed the AI’s work and assumes full responsibility for the results.