ºÚÁÏÊÓƵ

SoC Seminar presented by PhD student, Ulugbek Ergashev, April 25, 2025 from Noon - 1 in EB T1

°Õ¾±³Ù±ô±ð:ÌýResource selection in federated search

´¡²ú²õ³Ù°ù²¹³¦³Ù:ÌýIn today’s digital landscape, users often face challenges in accessing relevant information scattered across multiple, disparate sources. Federated Search (also known as Distributed Information Retrieval, DIR) offers a promising solution by enabling the retrieval of integrated results from heterogeneous repositories without the need for centralized indexing. However, existing resource selection methods in Federated Search are limited in capturing the complex semantic relationships between queries and resources. This research proposes novel approaches to enhance resource selection by leveraging advanced representation learning techniques. The primary contributions include FedGNN, a framework that utilizes Graph Neural Networks (GNNs) combined with pre-trained language models to capture both semantic and structural relationships between queries and resources, and Resource2Box, a method that models resources as hyperrectangular boxes in a latent space, enabling a richer representation of resource diversity. These methods aim to improve the accuracy and efficiency of resource ranking by integrating both query-resource and resource-resource relationships. Experimental results demonstrate that the proposed methods significantly outperform existing baseline models in various Federated Search tasks. This work contributes to the scalability and adaptability of Federated Search systems, making them more effective in dynamic and decentralized environments.

µþ¾±´Ç:ÌýUlugbek Ergashev is a Ph.D. student in the School of Computing at ºÚÁÏÊÓƵ. He earned his M.S. in Computer Science from ºÚÁÏÊÓƵ in 2019. His current research focuses on improving resource selection in Federated Search systems, particularly in decentralized environments where data is dispersed across multiple, heterogeneous sources. His work addresses the limitations of traditional search systems by leveraging advanced representation learning techniques, such as Graph Neural Networks (GNNs) and box embeddings, to enhance resource ranking accuracy.

Zoom Link:ÌýÌý

Title: Exploration of 3D Robot Vision Across Multiple Image Modalities

Abstract: 3D robot vision across multiple image modalities has emerged as a powerful tool for enhancing perception and scene understanding in diverse applications. By integrating data from visible images, thermal imaging, LiDAR, and ground-penetrating radar (GPR), multi-modal 3D vision enables robust and comprehensive environmental mapping, object detection, and depth estimation. Each modality provides complementary information. For example, visible images offering rich textures, thermal imaging enhancing detection in low-light conditions, and LiDAR captures precise geometric structures. These capabilities drive advancements in robotics, autonomous driving, mobile computing, and plant science, enabling autonomous navigation, structural assessment, and environmental monitoring in complex and dynamic environments. This work explores the challenges and opportunities in fusing multi-modal data for 3D vision, addressing sensor alignment, data fusion strategies, and application-specific optimizations to improve real-world performance across domains.

Bio: Dr. Lu is an Associate Professor at SUNY ºÚÁÏÊÓƵ. Before joining Binghamton, he was an Assistant Professor at the University of Georgia and Rochester Institute of Technology (RIT), a Research Scientist in autonomous driving at Ford, and a Research Engineer at Disney ESPN Advanced Technology Group. He has served as Principal Investigator (PI) for projects funded by NSF, USDA, DoD, the Georgia Department of Agriculture, Ford, GM, the Georgia Peanut Foundation, Qualcomm, Tencent, Mackinac, and others. His contributions have been recognized with several prestigious awards, including the NSF CAREER Award, USDA New Investigator Award, Aharon Katzir Young Investigator Award from the International Neural Network Society (INNS), Ford URP Award, Tencent Rhino-Bird Young Faculty Award, Frank A. Pehrson Award, and Erasmus Mundus Scholarship. He serves as the Chair of the IEEE Atlanta Signal Processing Society Chapter and the Co-Chair of the IEEE Robotics & Automation Society (RAS) Technical Committee on Agricultural Robotics and Automation. His research focuses on robotic perception, computer vision, and deep/machine learning.