CEE Seminars and Conferences

View Full Calendar

PhD Final Defense – Amir Kazemi

Event Type
Conference/Workshop
Sponsor
Civil and Environmental Engineering
Location
Newmark Civil Engineering Laboratory Yeh Center Room 1311
Date
Nov 15, 2024   11:00 am  
Views
79

Advancing Optimal Transport and Denoising Diffusion in Generative AI for

Learning in Urban Monitoring Systems

Advisors: Professor Christopher Tessum and Professor Volodymyr Kindratenko

Abstract:

Generative AI, notably through optimal transport and denoising diffusion models (DDM), is

enhancing the synthesis and analysis of data in urban monitoring systems. This technology is

crucial for creating high-fidelity simulations and models that predict dynamics, supporting the

development of resilient infrastructures and optimized processes in the face of operational and

environmental challenges.

Optimal transport theory plays a pivotal role in learning traffic data distributions and enhancing

urban surveillance capabilities. The novel Iterative Generative Adversarial Networks for

Imputation (IGANI) architecture introduces a new approach to data completion using an iterative

GAN leveraged by optimal transport. Additionally, optimal transport theory is applied to

rigorously evaluate one-shot generative models for data augmentation, focusing on

radiofrequency (RF) signal patterns in proliferating Unmanned Aerial Vehicles (UAV). Such

one-shot generative distribution matching enhances RF-based UAV identification in data-limited

environments, enabling more reliable urban surveillance.

DDMs have also emerged as powerful generative tools in computer vision and scientific

machine learning. Prompt-guided outpainting DDMs are explored for generating contextually

rich images for traffic monitoring and expanding datasets with automatic annotations.

AI-generated Dataset of Outpainted Vehicles for Eye-level Classification and Localization

(AIDOVECL) reduces the need for manual labeling and improves performance metrics for

vehicle detection. Moreover, a novel approach has been proposed to capture the dynamics of

physical processes using DDMs. This new algorithm for DDM training and sampling

demonstrates its capability to simulate phenomena such as heat distribution, presenting

potential for enhancing efficient environmental planning and management.

Overall, this dissertation illustrates how generative AI, particularly through optimal transport and

DDMs, may address data scarcity and improve the fidelity and utility of synthetic data in urban

monitoring. The implementations presented highlight potentials in data-driven decision-making

within infrastructure systems for real-world scenarios. This work contributes to insights and

methodologies that could inspire further research and practical applications in urban monitoring

and beyond.

link for robots only