Uploaded May 2025 | Updated September 2026, 6 hours ago
Abstract: Generative modeling and representation learning are fundamental to modern machine learning and computer vision. A high-quality generative model that produces realistic images and videos must capture complex visual structures and patterns. This necessity establishes an intrinsic connection between generative modeling and representation learning: generative models develop rich internal representations to capture high-dimensional data distributions effectively. Conversely, understanding the internal mechanisms of generative models through representation learning provides insights into improving them. This talk will focus on the bidirectional connection. Specifically, I will explore two core questions: (1) How can we improve the representation learning capabilities of generative models? (2) Can we leverage representations to enhance generative modeling? By addressing these questions, I aim to highlight the necessity of designing, analyzing, and advancing generative models through the lens of representation learning.
Bio: Xiao Zhang is a final-year CS Ph.D student at the University of Chicago, where he primarily works with Michael Maire. His research focuses on computer vision, with an emphasis on unsupervised representation learning and generative models. He aims to design scalable algorithms for representation learning that can be applied to real-world tasks and complex image and video generation.
Abstract: Generative modeling and representation learning are fundamental to modern machine learning and computer vision. A high-quality generative model that produces realistic images and videos must capture complex visual structures and patterns. This necessity establishes an intrinsic connection between generative modeling and representation learning: generative models develop rich internal representations to capture high-dimensional data distributions effectively. Conversely, understanding the internal mechanisms of generative models through representation learning provides insights into improving them. This talk will focus on the bidirectional connection. Specifically, I will explore two core questions: (1) How can we improve the representation learning capabilities of generative models? (2) Can we leverage representations to enhance generative modeling? By addressing these questions, I aim to highlight the necessity of designing, analyzing, and advancing generative models through the lens of representation learning.
Bio: Xiao Zhang is a final-year CS Ph.D student at the University of Chicago, where he primarily works with Michael Maire. His research focuses on computer vision, with an emphasis on unsupervised representation learning and generative models. He aims to design scalable algorithms for representation learning that can be applied to real-world tasks and complex image and video generation.
