Date of Award

Fall 1-1-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Statistics and Data Science

First Advisor

Yildirim, Ilker

Abstract

This dissertation investigates the evolving nature of model generalization in machine learning, with a particular focus on out-of-distribution behavior. It begins by examining how feature selection identifies and discards non-generalizable features, then explores the robustness of ImageNet-based models under shifted data distributions and their transfer to downstream tasks. The discussion moves to Vision-Language Models, exposing a "concept association bias'' that stems from contrastive training and exploring why autoregressive methods can offer stronger relational understanding. It then discuss how 3D-based pre-training mitigates spurious background-object correlations, improving generalization. Finally, it examines Large Language Models' spatial reasoning abilities, both in 2D and 3D, revealing how text-based systems can, with the aid of external tools, design novel 3D objects and navigate spatial tasks beyond their training distribution.

Download

Share

COinS