DS Ph.D. Qualifier Presentation | Nicholas Josselyn | Thurs. Dec. 7th @ 12:00pm EST
12:00 p.m. to 1:00 p.m.
DATA SCIENCE
Ph.D. Qualifier Presentation
Nicholas Josselyn
Thursday, December 7th , 2023 | 12:00 PM EST
Location: Gordon Library 303 Conference Room
Committee:
Prof. Elke Rundensteiner, research advisor
Prof. Ziming Zhang, research advisor
Prof. Randy Paffenroth, Qual co-advisor
Prof. Reza Zekavat, Qual co-advisor
Title:
Ensembles of Statistically Independent Models: An Embarrassingly Simple yet Strong Method for Domain Adaptation
Abstract:
Over recent years many sophisticated methods for the critical problem of domain adaptation (DA) have been proposed in literature. Unfortunately, there is no single DA model that always outperforms others regardless of datasets and setups. Motivated by this, we propose to investigate the effect of ensembles of (imperfect) classic DA models each pre-trained on distinct folds of data on the ability to learn a simple yet strong DA model for solving a particular new target problem. We find that a simple fusion of approximately statistically independent models can significantly boost the performance over each individual DA model. We demonstrate that, when taken together, the rich variety of these ensembled models succeeds to better cover the feature space for domain adaptation, achieving state-of-the-art performance on benchmark data sets for both single-source adaptation like DomainNet (+3.5%) and multi-source adaptation for Office-31 and DomainNet (+3.1% and +5.9%, respectively).
Research was sponsored by DEVCOM Army Research Laboratory under Cooperative Agreement W911NF-19-2-0112, W911NF-17-2-0227, W911NF1920108, and partially supported by DOE GANN P200A180088 Fellowship and NSF grants 1815866, 1910880, CCF-2006738, NRT-CEDAR 2021871.