More about HKUST
From Data Scaling to Data Intelligence: Data-Centric Learning for Generalizable Multimodal Foundation Models
PhD Qualifying Examination
Title: "From Data Scaling to Data Intelligence: Data-Centric Learning for
Generalizable Multimodal Foundation Models"
by
Mr. Haokun GUI
Abstract:
Multimodal foundation models increasingly depend on heterogeneous data
spanning vision, language, audio, embodied trajectories, and model-generated
interactions. As these systems expand beyond web-scale supervision, data
become costly, unevenly distributed, and difficult to scale, making data
utility as important as data volume.
This survey develops a data-centric framework for multimodal foundation models
across the data lifecycle. It organizes existing approaches around data
construction, cleaning and alignment, representation learning, adaptive
information selection, and feedback-driven improvement. The survey examines
how these mechanisms improve coverage, transferability, sample efficiency, and
generalization under limited data and compute budgets. Emerging evidence
suggests a shift from data scaling toward data intelligence, where models
increasingly optimize which data to collect, how to structure and exploit
them, and how to convert inference-time interactions and failures into new
supervision. Closed-loop data generation, valuation, and reuse remain central
research challenges.
Date: Monday, 24 August 2026
Time: 11:00am - 1:00pm
Venue: Room 2408
Lift 17/18
Committee Members: Prof. Jiaya Jia (Supervisor)
Dr. Dan Xu (Chairperson)
Dr. May Fung