Amit Kohli

EECS Department, University of California, Berkeley

Technical Report No. UCB/

December 1, 2025

Computational imaging has transformed scientific measurement by co-designing optical hardware and reconstruction algorithms around mathematical models of image formation. However, as applications push toward increasingly extreme scales (e.g., endoscopy), more demanding specifications in resolution and field of view (e.g., whole brain imaging), or the extraction of more information from fewer measurements (e.g., snapshot spectral imaging), computational imaging’s lack of robustness becomes apparent. In such settings, conditions deviate from idealized assumptions, and computational imaging systems can fail catastrophically or produce misleading results.

Among the prominent sources of error are optical aberrations and algorithmic uncertainty. Aberrations—imperfections present in all real imaging systems—are often inadequately modeled and can be computationally irreversible, limiting what can be recovered through standard post-processing alone. Algorithmic uncertainty arises when reconstruction algorithms lack guarantees on their behavior and may produce errors that are unexpectedly large or appear plausible but are incorrect. Deep learning models are an example of such algorithms; their flexibility makes them attractive for challenging reconstruction tasks, but their blackbox nature stymies interpretability and reliability.

This dissertation develops principled methods to address these challenges across the computational imaging pipeline. First, we introduce ring deconvolution microscopy, exploiting rotational symmetry to efficiently correct spatially-varying aberrations. Second, we prove that incorporating random phase masks into optical systems dramatically reduces their sensitivity to unknown aberrations, making their effects computationally reversible. Finally, we develop statistically rigorous uncertainty quantification for deep learning-based reconstruction, providing pixel-wise confidence intervals with formal guarantees that reveal unreliable regions. Together, these contributions establish foundations for robust computational imaging under real-world conditions.

Advisors: Laura Waller

---

BibTeX citation:

@phdthesis{Kohli:31994,
    Author= {Kohli, Amit},
    Title= {Robust Computational Imaging Under Aberrations and Algorithmic Uncertainty},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Month= {Dec},
    Number= {UCB/},
    Abstract= {Computational imaging has transformed scientific measurement by co-designing optical hardware and reconstruction algorithms around mathematical models of image formation. However, as applications push toward increasingly extreme scales (e.g., endoscopy), more demanding specifications in resolution and field of view (e.g., whole brain imaging), or the extraction of more information from fewer measurements (e.g., snapshot spectral imaging), computational imaging’s lack of robustness becomes apparent. In such settings, conditions deviate from idealized assumptions, and computational imaging systems can fail catastrophically or produce misleading results.

Among the prominent sources of error are optical aberrations and algorithmic uncertainty. Aberrations—imperfections present in all real imaging systems—are often inadequately modeled and can be computationally irreversible, limiting what can be recovered through standard post-processing alone. Algorithmic uncertainty arises when reconstruction algorithms lack guarantees on their behavior and may produce errors that are unexpectedly large or appear plausible but are incorrect. Deep learning models are an example of such algorithms; their flexibility makes them attractive for challenging reconstruction tasks, but their blackbox nature stymies interpretability and reliability.

This dissertation develops principled methods to address these challenges across the computational imaging pipeline. First, we introduce ring deconvolution microscopy, exploiting rotational symmetry to efficiently correct spatially-varying aberrations. Second, we prove that incorporating random phase masks into optical systems dramatically reduces their sensitivity to unknown aberrations, making their effects computationally reversible. Finally, we develop statistically rigorous uncertainty quantification for deep learning-based reconstruction, providing pixel-wise confidence intervals with formal guarantees that reveal unreliable regions. Together, these contributions establish foundations for robust computational imaging
under real-world conditions.},
}

EndNote citation:

%0 Thesis
%A Kohli, Amit 
%T Robust Computational Imaging Under Aberrations and Algorithmic Uncertainty
%I EECS Department, University of California, Berkeley
%D 2025
%8 December 1
%@ UCB/
%F Kohli:31994