Improved robustness of deep learning models through posterior agreement based model selection
Year: 2024
Type: dissertation
Abstract: Posterior Agreement (PA) has been proposed as a theoretically-grounded alternative for model robustness assessment in covariate shift settings. In this work, we provide further evidence in favor of these findings, and we explore the use of PA as a model selection criterion for deep learning models in supervised image classification tasks. Starting from the theoretical principles leading to the formulation of PA, we derive an operative version for discrete hypothesis set problems and use it as a robustness metric in the presence of adversarially-perturbed and domain-shifted data. We show that PA outperforms standard accuracy-based metrics in both settings and possesses superior discriminative power and consistency across increasing levels of covariate shift.
Author: Jiménez Rodríguez, Víctor
Institutions: ETH Zurich, Institute for Machine Learning
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7048035818
