-
PhD Thesis Defense - Yunhao Ge
Thu, Nov 30, 2023 @ 09:00 AM - 11:00 AM
Thomas Lord Department of Computer Science
University Calendar
PhD Thesis Defense - Yunhao Ge
Committee Members: Laurent Itti (chair), Yan Liu, Greg Ver Steeg, Nicolas Schweighofer
Title: Learning Controllable Data Generation for Scalable Model Training
Abstract: As machine learning models grow in complexity and power, the demands on training datasets surge correspondingly, necessitating both greater volume and enhanced quality. Harnessing real data, however, brings to the fore several challenges, including the hefty costs and sluggishness of human annotations—particularly in the fields of vision and robotics. Further obstacles include biases, spurious correlations, privacy concerns, and copyright constraints.In this talk, I will explore the potential of controllable automatic data generators as a solution to these data-related challenges. We will delve into harnessing learning techniques to control different data generation properties, culminating in photorealistic quality and significantly enhancing the training and performance of downstream models. Key insights include: ·
Methods to learn control over varying attributes, categories, distributions, and physical properties to bolster both 2D and 3D model training.
The transition of control from humans to downstream models, and how it paves the way for on-demand data generation, forging a symbiotic loop between the data generator and the downstream models.
A look ahead: The promise and challenges of generating intricate 3D and video data, underpinned by vision-language foundation models. We chart the frontier of controllable data generation and explore its vast potential in shaping the future of scalable model training.
Zoom Meeting ID: 222 662 0525Location: Hedco Neurosciences Building (HNB) - B15
Audiences: Everyone Is Invited
Contact: Melissa Ochoa
