Yanzhao Wu

Assistant Professor, KFSCIS


Lecture Information:
  • October 13, 2023
  • 2:00 PM
  • CASE 241

Speaker Bio

Dr. Yanzhao Wu is an Assistant Professor in the Knight Foundation School of Computing and Information Sciences at Florida International University. He obtained his bachelor’s degree from University of Science and Technology of China in 2017 and then received his Ph.D. degree in Computer Science from Georgia Institute of Technology in 2022. His research interests are primarily centered on the intersection of machine learning and computing systems, including machine learning algorithm and system optimizations, deep learning, large language models, edge AI, big data analytics, and their real-world applications. His work has been published in top venues, including CVPR, ICSE, IEEE ICDCS, IEEE ICDM, IEEE TPDS, IEEE TSC, and ACM TOIS, and won the IEEE CIC 2021 Best Paper Award. He also serves as a committee member/reviewer for top conferences and journals, such as ICDE, WWW, CVPR, ECCV, AAAI, IEEE TIFS, IEEE TKDE, and ACM TOIT.

Abstract

Deep neural network ensembles hold the potential to improve generalization performance and robustness for complex learning tasks. However, it remains an open challenge to harness model learning heterogeneity among member models to improve ensemble performance. This talk will present a two-tier heterogeneity driven ensemble framework powered by focal diversity to enhance the resilience and efficiency of ensemble learning systems. First, I will show that heterogeneous DNN models trained for solving the same learning problem, e.g., image classification or object detection, with high focal diversity, can significantly improve the generalization performance. Second, the ensemble robustness can be further strengthened by composing ensembles of heterogeneous models for solving different learning problems, e.g., object detection and semantic segmentation, through the proposed connected component labeling (CCL) based alignment. Third, I will introduce the focal diversity based ensemble pruning method to effectively identify high quality ensembles with high efficiency and high accuracy. Extensive experiments show that this focal diversity based two-tier heterogeneity driven ensemble framework can effectively leverage model learning heterogeneity to consistently boost ensemble robustness and efficiency. I will also provide an overview of our recent studies in Edge AI and Large Language Models (LLMs) and future research in machine learning systems.