Location
MC 5501
Speaker
Amine Aboussalah, New York University
Title
From Data Scarcity to Decision Resilience: A Principled Framework for Generative Modeling
Abstract
AI systems increasingly inform high-stakes decisions. Yet many machine learning methods are developed under idealized assumptions: abundant, clean, and stationary data. In many real-world settings, however, data are scarce, noisy, and continually shifting, which can lead to fragile models that fail precisely when robustness is most needed. This talk introduces a principled framework for decision-resilient modeling based on structured generative data augmentation. Rather than viewing augmentation as a purely heuristic device, we formulate it within a mathematical framework designed for structured data domains, including time series and graph-structured data, with explicit control of generalization under distributional shift. The analysis draws on statistical learning theory, in particular Rademacher-complexity bounds, to characterize out-of-sample performance and clarify the trade-offs induced by augmentation. We complement these guarantees with empirical results showing improved robustness, stability, and computational efficiency relative to existing generative methods. More broadly, this work positions generative modeling as a mathematically grounded layer for robust decision-making under uncertainty, offering a systematic pathway to bridge the simulation-to-reality gap in complex real-world applications.
Bio
Amine Mohamed Aboussalah is an Assistant Professor in the Department of Finance and Risk Engineering at the NYU Tandon School of Engineering. He received his PhD from the University of Toronto. His research seeks to understand the foundations of learning and dynamical systems through the lens of information geometry. He develops mathematically grounded frameworks that connect machine learning, reinforcement learning, and dynamical systems, with the goal of understanding the geometric principles that govern learning, adaptation, and decision-making in complex environments. He leads the Quantum Geometric Intelligence (QGI) Lab, which was established to further develop these connections and advance research at the interface of geometry, intelligence, and dynamical systems.