Ritwik Gupta

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2025-143

July 28, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-143.pdf

Perception in chaotic and uncertain environments, such as those marred by disasters and wars, remains a fundamental challenge in artificial intelligence. Characterized by data that is simultaneously partially observable, noisy, multi-spectral, extremely large, and with constantly shifting data biases, these environments undermine the assumptions on which most modern vision systems are built. This dissertation introduces a series of methodological contributions to address these challenges, including novel frameworks for representation learning under variable resolution, efficient processing of gigapixel imagery, and sequence modeling in high-dimensional visual domains. Alongside these technical advances, the work examines the governance implications of dual-use AI systems, proposing data-centric approaches to regulation and open-source methods for capability assessment. Together, these contributions aim to inform both the design and oversight of AI systems deployed in operationally complex and high-stakes settings.

Advisors: S. Shankar Sastry and Trevor Darrell

---

BibTeX citation:

@phdthesis{Gupta:EECS-2025-143,
    Author= {Gupta, Ritwik},
    Title= {Perception and Reasoning in Chaotic and Uncertain Environments},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Month= {Jul},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-143.html},
    Number= {UCB/EECS-2025-143},
    Abstract= {Perception in chaotic and uncertain environments, such as those marred by disasters and wars, remains a fundamental challenge in artificial intelligence. Characterized by data that is simultaneously partially observable, noisy, multi-spectral, extremely large, and with constantly shifting data biases, these environments undermine the assumptions on which most modern vision systems are built. This dissertation introduces a series of methodological contributions to address these challenges, including novel frameworks for representation learning under variable resolution, efficient processing of gigapixel imagery, and sequence modeling in high-dimensional visual domains. Alongside these technical advances, the work examines the governance implications of dual-use AI systems, proposing data-centric approaches to regulation and open-source methods for capability assessment. Together, these contributions aim to inform both the design and oversight of AI systems deployed in operationally complex and high-stakes settings.},
}

EndNote citation:

%0 Thesis
%A Gupta, Ritwik 
%T Perception and Reasoning in Chaotic and Uncertain Environments
%I EECS Department, University of California, Berkeley
%D 2025
%8 July 28
%@ UCB/EECS-2025-143
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-143.html
%F Gupta:EECS-2025-143